{"content_fingerprint":"3d2357202ce7519e7015df629f0c29a3670f89409a8994430443d0c8db8a5323","content":{"schema_version":2,"title":"VTX Insights Agent Operating Guide","eyebrow":"VTX Insights","hero":{"headline":"VTX Insights","summary":"Connect Codex, Claude Code, OpenCode, OpenClaw, Hermes Agent, Cursor, GitHub Copilot, Google Antigravity, or another compatible harness to power VTX bot inference or perform deep trade-chain analysis.","primary_action":"Open machine-readable guide","secondary_action":"Connect a harness"},"trust":{"agency":"Your agent decides how to investigate, calculate, recommend, and respond. VTX supplies the tools; it never grades or rewrites the answer.","permissions":"Choose analysis, settings management, bot control, profile management, connection management, and trading access separately. Your agent app receives only the permissions you approve.","ownership":"Changes and trades are limited to profiles you own and still pass the same VTX trading safeguards.","product_boundary":"Insights is built for trading and bot operations. Manage billing and account security directly in VTX."},"sections":[{"id":"overview","title":"Trade-chain analysis","summary":"Connect the available settings, decisions, executions, and outcomes behind a bot's trading campaigns.","blocks":[{"heading":"Reconstruct the available chain","body":"Connect enabled settings generations and decision-time context to model reasoning, execution attempts, fills, position campaigns, exits, fees, funding, and realized results. Whole-platform questions use one complete background calculation with focused retrieval of the compact evidence needed."},{"heading":"Explain the mechanism","body":"Separate strategy mistakes from market variance, thin liquidity, timing, overtrading, sizing, exit management, execution quality, and recent configuration changes."},{"heading":"Improve from evidence","body":"Compare losing chains with profitable counterexamples, preserve what works, and distinguish retained, same-decision repaired, recomputed, or unavailable evidence before proposing a small reversible experiment."}]},{"id":"connect","title":"Connect a harness","summary":"Choose a compatible harness, install the VTX connector, and approve only the access it needs.","steps":["Choose Codex or Claude Code for a bundled plugin, or connect OpenCode, OpenClaw, Hermes Agent, or another compatible agent app and install the same VTX skills separately.","Use the agent's quick install or the manual setup shown on this page.","Sign in to VTX and approve the analysis or action permissions you want this agent to use.","Ask the agent to find your profiles, analyze a question, or propose an optimization before making a change."]},{"id":"agent-guide","title":"Agent operating guide","summary":"Start with the machine-readable contract, then use the live capability registry to analyze evidence, operate bots, and verify the result.","blocks":[{"heading":"Power bot inference","body":"Use the connected harness as the live inference engine behind Main, Review, or Screener. VTX supplies the exact prompt, market context, and output contract; the harness returns the model response or structured decision that powers the normal VTX trading loop."},{"heading":"Analyze the trade chain","body":"Authentication identifies the VTX account using the connector, but it does not establish whether a supplied profile handle is owned or public. For named-profile work, discover the profiles once, confirm the ownership label for every requested handle or wallet, use the matching exact owned or public selection, and preserve that resolved selection with one evidence cutoff. Then connect effective settings and prompt generations to decisions, reasoning, executions, fills, position campaigns, fees, funding, and realized outcomes. Report unavailable evidence instead of silently narrowing the claim."},{"heading":"Run and verify a bot","body":"Resolve the owned profile, read its current execution mode and runtime state, request bot-control access, and start the normal VTX Trader. Server Mode runs on VTX and uses the normal fee. A Client Mode start records the requested intent but still needs a fresh owner: open that profile on Trade in a browser or desktop app, or run vtx auth login --scopes read,bot:control followed by vtx --profile runtime run --follow for the supported headless path. The bot is not running until current heartbeat and ownership evidence confirm it."},{"heading":"Use the capability registry","body":"Search the live capability registry instead of guessing tool names. Read each capability's permission and input contract, preview consequential changes, ask for approval, and verify receipts against effective state and observed results. A bot is not running until current runtime evidence confirms it."}],"steps":["Read https://api.vtxmacro.com/insights/access-center/metadata, choose the appropriate bundled skill and live capability, request only the permissions required for my outcome, and explain what you will verify before taking any consequential action."]},{"id":"skills","title":"Bundled skills","summary":"Codex and Claude Code plugin installs include both skills. OpenCode, OpenClaw, and Hermes Agent can install the same skills separately, and other compatible harnesses can use the VTX capabilities through MCP.","blocks":[{"heading":"vtx-insights-analysis","body":"The general VTX router and safety contract. It discovers the right capability, preserves complete evidence, separates permissions, and verifies settings, profile, connection, bot-control, and trading actions."},{"heading":"vtx-bot-trade-chain-analysis","body":"The specialist performance workflow. It reconstructs settings generations, decisions, executions, fills, positions, exits, fees, funding, PnL, and market context before recommending a focused improvement."}],"steps":["Use $vtx-insights-analysis to work with VTX safely and choose the right capability.","Use $vtx-bot-trade-chain-analysis to explain why my bots are winning or losing and recommend the smallest evidence-backed improvements."]},{"id":"use","title":"Use","summary":"Describe the outcome you want and let the agent choose the best VTX and native tools for the job.","steps":["Name the population: my profiles, exact public handles or wallets, or the whole platform. For named profiles, the agent verifies each ownership label before choosing the owned or public selection.","Name the time window and the decision you are trying to make.","Ask for the analysis, comparison, recommendation, or action that helps you decide.","Review important assumptions and approve consequential changes or trades in plain language.","Treat policy replay as a counterfactual, not as observed live performance."]},{"id":"examples","title":"Examples","summary":"Review examples drawn from current VTX trading activity, then copy and adapt a prompt.","blocks":[{"heading":"Adapt the prompt","body":"Use the copyable prompt as a starting point, change the profiles or time window, and ask your agent to investigate what matters to you."},{"heading":"Check the as-of time","body":"Live and recently captured figures can change as bots trade and new records arrive. Compare the displayed as-of time before treating two runs as contradictory."}]},{"id":"prompt-ideas","title":"Prompt ideas","summary":"Search starting points for performance reviews, trade-chain analysis, settings ideas, market research, and VTX product help.","blocks":[{"heading":"Prompts are starting points","body":"They improve discovery but do not prescribe tool sequences or an answer template. Add the population, time window, comparison dimensions, and verification standard that matter to you."}]},{"id":"capabilities","title":"Capabilities","summary":"Browse capabilities by permission group, or search for an exact analysis or owned-profile action.","blocks":[{"heading":"Population and coverage handling","body":"VTX tools preserve the selected population, resume large results from acknowledged checkpoints, and report unavailable evidence instead of silently substituting a smaller answer."}]},{"id":"troubleshooting","title":"Troubleshooting","summary":"Get a connection or workflow back on track quickly.","items":[{"question":"My agent app cannot find VTX Insights","answer":"Refresh its plugin or MCP integrations, confirm VTX Insights is enabled, or use the manual setup shown on this page. Update the app or integration if it is still unavailable."},{"question":"VTX asks me to sign in again","answer":"Start the agent's VTX sign-in flow, choose the permissions you want, and approve the connection. If access was revoked or expired, relink the client from this page."},{"question":"My agent reports Auth required","answer":"Start VTX authentication from the MCP client and approve the browser request. In ChatGPT desktop use Settings, MCP servers, VTX Insights, and Authenticate. In Codex CLI or the IDE run codex mcp login vtx-insights --scopes insights:read."},{"question":"Codex reports Authorization server response missing required issuer","answer":"Codex 0.146.0 has a known OAuth issuer-validation regression. Check codex --version, upgrade Codex, and start a new VTX login."},{"question":"My agent reports Insights tool discovery failed","answer":"OAuth succeeded, but VTX could not return the authorized MCP tool catalog. This is a VTX MCP service or catalog problem, not a request for more browser or extension permissions. Retry after VTX service health is restored."},{"question":"VTX reports Authorization request is unavailable","answer":"That authorization link is no longer usable. Return to the MCP client and start a new login; do not reuse the URL."},{"question":"A result looks incomplete","answer":"Ask which profiles and time window were used, what was unavailable, and whether population coverage differs from exact-economics coverage. Replay and excursion calculations run when requested; they are not an always-running simulator or an exact alternate-PnL forecast."},{"question":"My agent reports profile selection unavailable","answer":"At least one exact handle or wallet could not be resolved in the selected owned or public population. Ask the agent to discover the profiles again, confirm each ownership label, and retry with the corrected exact selection. VTX stops before starting a background analysis rather than silently analyzing a smaller population."},{"question":"The model's conclusion looks wrong","answer":"Ask it to show the VTX evidence and calculations, challenge the assumptions, or try a different approach. Models can make mistakes; VTX does not approve or correct their conclusions."},{"question":"An analysis is taking a long time","answer":"The agent can continue the complete analysis in the background and resume retrieval from its last acknowledged checkpoint without shrinking the profiles or time window you asked for."},{"question":"My agent cannot query a targeted artifact path","answer":"Targeted discovery is recursive and does not change the server index. The agent should query the artifact root, then each needed disclosed object path and continuation, then the disclosed list parent before batching exact child paths. Large typed filters scan that exact indexed parent incrementally and may continue through background analysis if they exceed the response window."}]},{"id":"faq","title":"FAQ","summary":"The essentials of agent choice, permissions, trading, and control.","items":[{"question":"Can VTX Insights trade or change my bot?","answer":"Yes. With your approval, a connected agent can update AI settings on profiles you own, control Server Mode bots, request Client Mode Trader starts and stops, and place or cancel orders through VTX. Client control does not take over an owner; waiting for an eligible browser, desktop, or headless owner is not the same as running, and an in-flight stop may wait for settlement before its owner lease is revoked."},{"question":"Do I need to teach my agent how to analyze a trade chain?","answer":"No. The VTX Insights connector ships a dedicated trade-chain analysis skill for Codex, Claude Code, OpenCode, OpenClaw, Hermes Agent, and other compatible skill hosts. It teaches the agent how to reconstruct bot campaigns, compare settings generations, verify evidence, and recommend focused improvements."},{"question":"Can I analyze another public VTX profile?","answer":"Yes. Select exact public handles or wallets, or compare the whole platform. Actions always remain limited to profiles you own."},{"question":"Does VTX choose the model?","answer":"No. The connected agent app owns model choice, conversation, inference, and native tools."},{"question":"Are large results summarized?","answer":"The response may summarize the result, but complete evidence remains available for focused indexed queries or restart-safe acknowledged retrieval."},{"question":"Where should I ask general VTX support questions?","answer":"Use VTX Help for product guidance outside Insights, or join the VTX Macro Telegram group for community help and setup discussion. Connected agent apps can also search and open the same canonical Help content.","links":[{"label":"Open VTX Help","href":"/help"},{"label":"Join Telegram","href":"https://t.me/VTXMacro","external":true}]}]},{"id":"lifecycle","title":"Update, disconnect, and recover","summary":"Keep the agent integration current and remove access whenever a connected app is no longer trusted.","steps":["Refresh the agent app's marketplace or registry and install the current compatible package version.","Disable the integration temporarily when supported, or uninstall it from the agent app.","Disconnect the agent from this page whenever you no longer want it to use VTX.","Use Relink to remove the old connection first, then sign in again from the agent.","If an agent remains stuck, remove its saved VTX connection, reinstall or reconfigure it, and sign in again."]}]},"distribution":{"schema_version":2,"distribution_id":"vtx-insights","package_version":"2026.9.3","shipped_skill_agents":[{"path":"skills/vtx-bot-trade-chain-analysis/SKILL.md","sha256":"949ff62fbb329dbef71dd1a763d9a4c518980d72e4af8e6ad8bc5f67c4fb5874"},{"path":"skills/vtx-bot-trade-chain-analysis/agents/openai.yaml","sha256":"25d87c5bf11f3fee92b14d3de50bc244ea482712d61768df53730e509ac490f3"},{"path":"skills/vtx-insights-analysis/SKILL.md","sha256":"ec008b3a4284a8eecef96ebeda60f7cf21c3f7e5e42d5beca4cdb8348a9d7202"}],"publisher":{"name":"DataDoesYou LLC","repository":"https://github.com/DataDoesYou/VTX-Insights"},"service":{"name":"VTX Insights","website_url":"https://vtxmacro.com/insights","privacy_url":"https://vtxmacro.com/privacy","terms_url":"https://vtxmacro.com/terms","mcp_url":"https://api.vtxmacro.com/insights/mcp","oauth_scope":"insights:read","oauth_scopes":["insights:read","insights:settings","insights:control","insights:trade","insights:profiles","insights:credentials"],"preferred_mcp_protocol_version":"2026-07-28","supported_mcp_protocol_versions":["2026-07-28","2025-11-25"],"mcp_protocol_version":"2026-07-28","minimum_server_contract":1,"maximum_server_contract":1},"hosts":{"codex":{"display_name":"Codex","recommended":true,"distribution":"marketplace","listing_status":"published","marketplace_name":"vtx-insights","mcp_tool_timeout_seconds":120,"install":["codex plugin marketplace add DataDoesYou/VTX-Insights","codex plugin add vtx-insights@vtx-insights","codex mcp login vtx-insights --scopes insights:read"],"update":["codex plugin marketplace upgrade vtx-insights","codex plugin remove vtx-insights@vtx-insights","codex plugin add vtx-insights@vtx-insights"],"uninstall":["codex plugin remove vtx-insights@vtx-insights"],"manual_config":"[mcp_servers.vtx-insights]\nurl = \"https://api.vtxmacro.com/insights/mcp\"\ntool_timeout_sec = 120","auth_steps":["Authentication starts in the MCP client, not on this page. In ChatGPT desktop, open Settings → MCP servers, select VTX Insights, and choose Authenticate","In Codex CLI or the IDE, run codex mcp login vtx-insights --scopes insights:read and approve the browser request. Request additional permissions only when a later task needs them","Run codex --version before troubleshooting OAuth. Codex 0.146.0 has a known issuer-validation regression; upgrade it if login reports Authorization server response missing required issuer"],"verify_steps":["Restart ChatGPT desktop or open a new Codex CLI or IDE session after authentication so initialization and tool discovery run again","In ChatGPT desktop, type /mcp and confirm VTX Insights is active; in CLI or the IDE, inspect the new session's MCP tool inventory","Confirm expected read tools such as analysis.start and positions.episodes are available. An OAuth grant alone does not prove that initialization and tool discovery succeeded"]},"claude_code":{"display_name":"Claude Code","recommended":false,"distribution":"marketplace","listing_status":"published","marketplace_name":"vtx-insights","install":["claude plugin marketplace add DataDoesYou/VTX-Insights","claude plugin install vtx-insights@vtx-insights","claude mcp login plugin:vtx-insights:vtx-insights"],"update":["claude plugin marketplace update vtx-insights","claude plugin uninstall vtx-insights@vtx-insights","claude plugin install vtx-insights@vtx-insights"],"disable":["claude plugin disable vtx-insights@vtx-insights"],"enable":["claude plugin enable vtx-insights@vtx-insights"],"uninstall":["claude plugin uninstall vtx-insights@vtx-insights"],"manual_config":"{\n \"mcpServers\": {\n \"vtx-insights\": {\n \"type\": \"http\",\n \"url\": \"https://api.vtxmacro.com/insights/mcp\"\n }\n }\n}"},"cursor":{"display_name":"Cursor","recommended":false,"distribution":"public_marketplace","listing_status":"submission_pending","install":["/add-plugin vtx-insights"],"manual_config":"{\n \"mcpServers\": {\n \"vtx-insights\": {\n \"type\": \"http\",\n \"url\": \"https://api.vtxmacro.com/insights/mcp\"\n }\n }\n}"},"github_copilot":{"display_name":"GitHub Copilot","recommended":false,"distribution":"mcp_registry","listing_status":"publication_pending","registry_name":"io.github.DataDoesYou/vtx-insights","manual_config":"{\n \"servers\": {\n \"vtx-insights\": {\n \"url\": \"https://api.vtxmacro.com/insights/mcp\"\n }\n }\n}"},"google_antigravity":{"display_name":"Google Antigravity","recommended":false,"distribution":"mcp_store","listing_status":"unlisted","marketplace_name":"VTX Insights","install_format":"steps","install":["Open Settings → Customizations → Installed MCP Servers","Choose Add MCP, then Manage MCP Servers → View raw config","Paste the manual configuration below and save it"],"manual_config":"{\n \"mcpServers\": {\n \"vtx-insights\": {\n \"serverUrl\": \"https://api.vtxmacro.com/insights/mcp\"\n }\n }\n}","auth_steps":["Select Authenticate next to VTX Insights","Complete VTX sign-in in your browser, copy the authorization code, and paste it back into Antigravity","Select Submit and confirm the server reconnects"],"plugin_install":["Workspace: copy plugins/vtx-insights to .agents/plugins/vtx-insights (or _agents/plugins/vtx-insights)","Global: copy plugins/vtx-insights to ~/.gemini/config/plugins/vtx-insights"]},"opencode":{"display_name":"OpenCode","recommended":false,"distribution":"native_mcp","listing_status":"unlisted","install":["Save the manual configuration below as opencode.json in your project or global OpenCode config directory","opencode mcp auth vtx-insights","opencode mcp list"],"install_format":"steps","manual_config":"{\n \"$schema\": \"https://opencode.ai/config.json\",\n \"mcp\": {\n \"vtx-insights\": {\n \"type\": \"remote\",\n \"url\": \"https://api.vtxmacro.com/insights/mcp\",\n \"enabled\": true,\n \"oauth\": {\n \"scope\": \"insights:read\"\n },\n \"timeout\": 120000\n }\n }\n}","auth_steps":["Run opencode mcp auth vtx-insights and complete VTX sign-in in your browser","Approve only the VTX permissions this OpenCode session should be able to use","Run opencode mcp list to confirm VTX Insights is connected; use opencode mcp debug vtx-insights if authentication fails"],"skill_install":["git clone https://github.com/DataDoesYou/VTX-Insights.git","Copy VTX-Insights/plugins/vtx-insights/skills/vtx-insights-analysis to ~/.config/opencode/skills/vtx-insights-analysis","Copy VTX-Insights/plugins/vtx-insights/skills/vtx-bot-trade-chain-analysis to ~/.config/opencode/skills/vtx-bot-trade-chain-analysis"]},"openclaw":{"display_name":"OpenClaw","recommended":false,"distribution":"native_mcp","listing_status":"unlisted","install":["openclaw mcp add vtx-insights --url https://api.vtxmacro.com/insights/mcp --transport streamable-http --auth oauth --oauth-scope \"insights:read\" --timeout 120","openclaw mcp login vtx-insights","openclaw mcp login vtx-insights --code YOUR_AUTHORIZATION_CODE","openclaw mcp doctor vtx-insights --probe"],"manual_config":"{\n \"mcp\": {\n \"servers\": {\n \"vtx-insights\": {\n \"url\": \"https://api.vtxmacro.com/insights/mcp\",\n \"transport\": \"streamable-http\",\n \"requestTimeoutMs\": 120000,\n \"connectionTimeoutMs\": 5000,\n \"auth\": \"oauth\",\n \"oauth\": {\n \"scope\": \"insights:read\"\n }\n }\n }\n }\n}","auth_steps":["Open the authorization URL printed by openclaw mcp login vtx-insights and complete VTX sign-in","Approve only the VTX permissions this OpenClaw agent should be able to use","Copy the authorization code and pass it to openclaw mcp login vtx-insights --code YOUR_AUTHORIZATION_CODE","Confirm openclaw mcp doctor vtx-insights --probe lists the expected VTX tools"],"skill_install":["git clone https://github.com/DataDoesYou/VTX-Insights.git","openclaw skills install ./VTX-Insights/plugins/vtx-insights/skills/vtx-insights-analysis --global","openclaw skills install ./VTX-Insights/plugins/vtx-insights/skills/vtx-bot-trade-chain-analysis --global"]},"hermes":{"display_name":"Hermes Agent","recommended":false,"distribution":"native_mcp","listing_status":"unlisted","install":["hermes mcp add vtx-insights --url https://api.vtxmacro.com/insights/mcp --auth oauth --connect-timeout 315","hermes config set mcp_servers.vtx-insights.trust untrusted","hermes mcp test vtx-insights"],"manual_config":"{\n \"mcp_servers\": {\n \"vtx-insights\": {\n \"url\": \"https://api.vtxmacro.com/insights/mcp\",\n \"enabled\": true,\n \"timeout\": 120,\n \"connect_timeout\": 315,\n \"auth\": \"oauth\",\n \"trust\": \"untrusted\",\n \"oauth\": {\n \"scope\": \"insights:read\"\n }\n }\n }\n}","auth_steps":["Use Hermes Agent 0.20.4 or newer from the supported official installer, Desktop app, or Docker image; the unsupported PyPI package does not provide this safety contract","Complete VTX sign-in when Hermes opens the OAuth authorization flow","Approve only the VTX permissions this Hermes agent should be able to use","Keep the VTX server trust tier set to untrusted so write-capable tools require local approval","Confirm hermes mcp test vtx-insights lists the expected VTX tools"],"skill_install":["hermes skills install https://raw.githubusercontent.com/DataDoesYou/VTX-Insights/main/plugins/vtx-insights/skills/vtx-insights-analysis/SKILL.md","hermes skills install https://raw.githubusercontent.com/DataDoesYou/VTX-Insights/main/plugins/vtx-insights/skills/vtx-bot-trade-chain-analysis/SKILL.md"]}}},"subscription_inference":{"schema_version":1,"distribution_id":"vtx-external-inference-host","listing_status":"published","general_admission":"enabled","service":{"mcp_url":"https://api.vtxmacro.com/insights/mcp","oauth_scope":"insights:inference","oauth_scopes":["insights:inference"],"grant_class":"external_inference_runtime","operational_tools":["inference.agent.connect","inference.agent.heartbeat","inference.agent.next","inference.agent.complete","inference.agent.fail","inference.agent.assignment.next","inference.agent.assignment.heartbeat","inference.agent.data.call","inference.agent.decision.submit","inference.agent.decision.status","inference.agent.assignment.release","inference.host.register","inference.host.heartbeat","inference.host.advertise","inference.job.claim","inference.job.start","inference.job.heartbeat","inference.job.complete","inference.job.fail"]},"cli":{"package_name":"@vtxmacro/cli","public_registry_status":"published","accepted_distribution":"public_npm","commands":["vtx inference-host --help","vtx inference-host login","vtx inference-host agent-connect","vtx inference-host agent-run","vtx inference-host agent-next","vtx inference-host agent-complete","vtx inference-host agent-fail","vtx inference-host agent-assignment-next","vtx inference-host agent-assignment-heartbeat","vtx inference-host agent-data-call","vtx inference-host agent-decision-submit","vtx inference-host agent-decision-status","vtx inference-host agent-assignment-release","vtx inference-host codex-login","vtx inference-host grok-login","vtx inference-host deepseek-login","vtx inference-host pi-login","vtx inference-host exo-login","vtx inference-host run","vtx inference-host service install","vtx inference-host service start","vtx inference-host service stop","vtx inference-host service status","vtx inference-host service logs","vtx inference-host service recover","vtx inference-host service uninstall","vtx inference-host status","vtx inference-host doctor","vtx inference-host logout","vtx inference-host codex-logout","vtx inference-host grok-logout","vtx inference-host deepseek-logout","vtx inference-host pi-logout","vtx inference-host exo-logout","vtx inference-host revoke"]},"help":{"title":"Connect an agent inference host","summary":"Connect Codex, GitHub Copilot, DeepSeek Harness, Pi, Exo, or another compatible harness to a running VTX bot. Provider mode keeps the normal VTX analysis cycle unchanged. On an explicitly capable Main model, Agent mode assigns the bot to the agent, which chooses its own cadence, requests only the VTX data it needs, and submits the normal structured trading decision.","runtime_requirement":"Provider mode supports Main, Review, and Screener while the normal VTX bot and connected host are running. Agent mode is available only for explicitly capable Main models; Review and Screener remain Provider-only. VTX assigns the running Trader but does not schedule Agent prompts or preassemble required context. The agent may request assignment-scoped data or none, then submits the same structured decision VTX validates and executes normally. Both Server Mode and Client Mode support these control modes. Durable workers use a supported headless runtime, restart at user login, and reconnect after sleep or network loss on native Windows or Linux. WSL remains foreground-only.","agent_prompt":"Read https://vtxmacro.com/insights#subscription-inference and help me connect your supported VTX inference host. Install the durable service when your harness has a supported headless adapter; otherwise use the foreground host path and keep its agent loop open. For Main, help me choose Provider or Agent based on the model's explicit capabilities; keep Review and Screener in Provider mode. Save the setting and start the normal VTX Trader only after I approve. Confirm either that the OS service remains active independently of this chat or that the required foreground loop is still open. Do not advertise durability or Agent control that your harness cannot provide.","fleet_setup":{"title":"Fresh computer or existing fleet","summary":"Use this path when several existing profiles or durable inference hosts must move to a new computer. A new installation creates new host identities even when the vendor credential is unchanged, so every saved profile lane must be checked and any stale binding must be explicitly previewed and rebound.","steps":["Before changing anything, freeze the exact profile cohort and record each lane's current model, reasoning effort, control mode, execution mode, primary host, and ordered fallbacks. Record every computer's displayed host name, stable local --instance name, adapter, host ID, and intended model. Record authenticated account identity only when that adapter reports one.","Install the CLI from native Windows or native Linux. On each physical computer, run login, then the adapter-specific vendor step: codex-login for Codex, the official Copilot CLI login for Copilot, or deepseek-login through stdin for DeepSeek Harness. Run doctor and service install with the exact adapter and instance; one OS supervisor manages that computer's isolated workers. Each worker keeps its own VTX grant, credential boundary, and runtime state. Never copy inference-host credentials or state between computers. Codex and Copilot multiple-instance setups require distinct authenticated accounts. DeepSeek Harness currently permits only one worker for that adapter per computer because it does not advertise a stable provider account identity.","Compare each newly registered host ID with the exact host identity saved on every target lane. Matching account email is not proof of the same host; neither is a matching API key, adapter, or device name. Preview an owner-scoped update for only the frozen cohort, explicitly replace stale primary and fallback host bindings, and preserve model, reasoning effort, response mode, execution mode, control mode, symbols, prompts, and every unrelated setting.","Apply only the approved, version-fenced preview with its rollback snapshot. Read back every target lane and Host cascade, then confirm each named instance maps to the intended host ID and adapter and advertises the intended model and effort. When account identity is available, verify it too. Do not revoke the previous computer's grants until the new bindings and workers are verified.","Verify a fresh completed and accepted inference from every primary host; for full fleet acceptance, verify one from every target profile. A routed request, claim, start acknowledgement, running Trader, start receipt, or brief Live label does not prove inference completion. Inspect failures before making one controlled retry.","A durable service starts immediately and registers startup for that user at sign-in, not before Windows login. While the computer is off, asleep, signed out, or offline it cannot serve inference. At sign-in and periodically while running, a desired-running service checks for a verified CLI update. It finishes active Provider calls and Agent turns with heartbeats intact before switching releases and reconnecting automatically; if the registry is unavailable, it keeps the last verified compatible runtime. service stop disables update checks, automatic recovery, and login startup until service start re-enables them.","Fallback workers installed on the same computer protect against an individual provider credential or quota failure, not against that computer being unavailable. Reboot or sign out only with the user's approval; after signing back in, verify supervisor state, every instance and adapter, any reported account identity, fresh remote heartbeats, and a new completed inference."],"verification":["Exact effective primary and fallback bindings read back for every target lane.","One OS supervisor reports every intended named worker desired-running with the correct adapter and, when available, authenticated account.","Every new host has a fresh remote heartbeat and advertises the intended model and reasoning effort.","Each primary host completes and VTX accepts a fresh inference; full-fleet acceptance covers every target profile.","A post-login check proves automatic restart when reboot or sign-out verification was approved."],"reboot_behavior":"The service is per-user login startup, not a pre-login machine service. Nothing runs while the computer is powered off, and workers cannot serve while it is asleep, signed out, or offline. After the user signs in, instances whose desired state is running reconnect automatically. Workers on the same computer share that computer-level failure domain.","agent_prompt":"Read https://vtxmacro.com/insights#subscription-inference and set up or move my exact existing durable inference fleet to this computer. Before changes, list the exact profiles, named instances, adapters, primary and fallback order, model, reasoning, control mode, and execution mode. Use the adapter-specific vendor login or key-import step and one native OS supervisor with isolated workers. DeepSeek Harness is limited to one worker for that adapter per computer until it can advertise a stable provider account identity. Explicitly preview and rebind stale host identities, preserve all unrelated settings, and verify effective readback plus a fresh completed inference from each primary. Do not call routing, a start receipt, or running Trader state proof of inference completion. Explain user-login startup and computer-off behavior before any reboot."},"adapters":[{"adapter":"codex","label":"OpenAI Codex","support_level":"durable","contract_status":"stable","billing_route":"subscription","structured_output":true,"same_attempt_recovery":true,"setup":"vtx inference-host codex-login; vtx inference-host service install --adapter codex","note":"The existing pinned App Server path remains the reference implementation, including ChatGPT identity, live model and effort discovery, quota telemetry, cancellation, and exact crash recovery."},{"adapter":"copilot","label":"GitHub Copilot","support_level":"durable","contract_status":"preview","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Sign in with the official Copilot CLI, then run vtx inference-host service install --adapter copilot","note":"VTX uses GitHub's official SDK with the package-owned pinned platform runtime, authenticated GitHub login identity, exact enabled model and effort events, token usage, and cooperative cancellation. Auto-only catalogs fail closed because they cannot prove the effective model. Output text is schema-validated by VTX; uncertain interrupted calls are quarantined."},{"adapter":"deepseek-harness","label":"DeepSeek Harness","support_level":"durable","contract_status":"preview","billing_route":"api_billed","structured_output":false,"same_attempt_recovery":false,"setup":"Import a DeepSeek API key through stdin with vtx inference-host deepseek-login, then run vtx inference-host service install --adapter deepseek-harness","note":"VTX uses the exact-pinned official DeepSeek Harness runtime in both Provider and Main Agent modes. The key remains in the local VTX private credential boundary, Provider runs have no tools, Agent runs expose only the three assignment-scoped VTX tools, and a loopback receipt verifies the provider-returned model and response ID. VTX validates final JSON; interrupted ambiguous work is never blindly retried."},{"adapter":"pi","label":"Pi","support_level":"durable","contract_status":"preview","billing_route":"api_billed","structured_output":true,"same_attempt_recovery":false,"setup":"Import an OpenAI API key through stdin with vtx inference-host pi-login --provider openai, then run vtx inference-host doctor --adapter pi and vtx inference-host service install --adapter pi","note":"The pinned Pi SDK supports Provider mode and Main Agent mode through the existing controls. The initially qualified models use one separately stored OpenAI API key per instance and advertise exact provider/model identities from the release's configured qualification policy and SDK catalog. Provider runs have no tools; persistent Agent sessions expose only the three assignment-scoped VTX tools. VTX validates structured output and execution receipts. Uncertain interrupted model calls stop safely; Pi does not claim same-attempt recovery. Subscription OAuth and unqualified provider protocols are not enabled by this adapter."},{"adapter":"exo","label":"Exo","support_level":"foreground","contract_status":"preview","billing_route":"api_billed","structured_output":false,"same_attempt_recovery":false,"setup":"On Linux or WSL, import a Venice key with vtx inference-host exo-login --provider venice --instance exo-1 (use --provider openai for OpenAI), then keep vtx inference-host run --adapter exo --exo-root /path/to/exo --instance exo-1 open","note":"Exo means exoharness/exo. The VTX-owned bridge runs the pinned Exo custom harness automatically in Provider and Main Agent modes; no manual job transfer is needed. Each instance advertises only the release-qualified routes for its OpenAI or Venice credential. Venice currently offers Gemma 4 31B IT with effort None only. Provider runs receive complete VTX input with no tools; Agent runs expose only assignment-scoped VTX tools. VTX validates canonical JSON; native structured-output enforcement is not claimed. VTX OAuth stays outside Exo. This Linux/WSL foreground preview has no durable service or same-attempt recovery: uncertain interrupted work remains fenced. It does not add native Exo MCP support."},{"adapter":"antigravity","label":"Google Antigravity","support_level":"foreground","contract_status":"preview","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Use vtx inference-host agent-connect and keep the signed-in Antigravity loop open","note":"A live 1.1.13 Starter Quota acceptance exposed shell, file, browser, web, MCP, subagent, and messaging tools even under an isolated agent declaring tools: [], and the requested final JSON Schema was not enforced. VTX therefore keeps Antigravity in a monitored foreground loop."},{"adapter":"gemini-cli","label":"Gemini CLI","support_level":"foreground","contract_status":"restricted","billing_route":"mixed","structured_output":false,"same_attempt_recovery":false,"setup":"Use the foreground agent loop or VTX's normal Gemini API provider; do not install Gemini CLI as a subscription worker","note":"The current official individual OAuth flow rejects this client and directs users to Antigravity. API-key and Vertex routes are API-billed rather than a consumer-subscription subsidy, and the CLI does not enforce a final response schema."},{"adapter":"claude-code","label":"Claude Code","support_level":"foreground","contract_status":"restricted","billing_route":"subscription","structured_output":false,"same_attempt_recovery":false,"setup":"Use vtx inference-host agent-connect and keep the Claude Code agent loop open","note":"Claude Code has a strong official programmatic contract, but Anthropic currently says third-party developers may not route Free, Pro, or Max credentials on users' behalf without prior approval. VTX therefore keeps it on the user's foreground first-party loop."},{"adapter":"grok-build","label":"Grok Build","support_level":"durable","contract_status":"preview","billing_route":"subscription","structured_output":true,"same_attempt_recovery":false,"setup":"Sign in with vtx inference-host grok-login, then install a grok-build inference host","note":"Grok Build uses the pinned official runtime and a separate local OAuth session for Provider and Main Agent modes. Provider runs have no tools; Agent runs expose only the three assignment-scoped VTX tools. VTX validates structured output, model identity, and reported usage. Uncertain interrupted requests remain fenced; same-attempt recovery and Codex subscription quota telemetry are not supported."},{"adapter":"kiro","label":"Kiro CLI","support_level":"foreground","contract_status":"insufficient","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Use the foreground agent loop; do not install Kiro as a durable VTX worker yet","note":"Kiro API keys legitimately consume paid subscription credits, but its headless chat lacks a stable JSON result, effective-model, token-usage, and cancellation receipt."},{"adapter":"cursor","label":"Cursor","support_level":"foreground","contract_status":"preview","billing_route":"api_billed","structured_output":false,"same_attempt_recovery":false,"setup":"Use Cursor's foreground agent loop for subscription sessions","note":"Cursor's official SDK has strong lifecycle and usage telemetry, but SDK authentication is API-key and token-billed rather than a clear consumer-subscription subsidy."},{"adapter":"amp","label":"Amp","support_level":"foreground","contract_status":"preview","billing_route":"mixed","structured_output":false,"same_attempt_recovery":false,"setup":"Use Amp through agent-connect while managed model-mode identity remains unfrozen","note":"Amp offers subscriptions, credits, and linked ChatGPT capacity, but its selectable product is a managed mode rather than a stable underlying model identity."},{"adapter":"auggie","label":"Auggie","support_level":"foreground","contract_status":"preview","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Use Auggie through agent-connect until structured results and recovery are complete","note":"Auggie exposes account, models, credits, ACP cancellation, and sessions, but its SDK currently declares structured output unsupported and recovery remains incomplete."},{"adapter":"junie","label":"JetBrains Junie","support_level":"foreground","contract_status":"insufficient","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Use Junie through agent-connect while its effective model, effort, and usage receipts remain incomplete","note":"Junie has headless JSON, ACP, sessions, and subscription routing, but effort can fall back and its public machine contract does not yet provide the exact effective identity and usage evidence required for a durable VTX worker."},{"adapter":"warp-oz","label":"Warp Oz","support_level":"foreground","contract_status":"insufficient","billing_route":"subscription_credits","structured_output":false,"same_attempt_recovery":false,"setup":"Use Warp or Oz through agent-connect for foreground automation","note":"Warp and Oz can consume plan credits and run cloud agents, but they do not expose the exact effort, structured result, token accounting, and direct inference lifecycle required for a durable VTX provider."},{"adapter":"qwen-code","label":"Qwen Code","support_level":"foreground","contract_status":"restricted","billing_route":"mixed","structured_output":true,"same_attempt_recovery":false,"setup":"Keep Coding Plan sessions foreground-only; use VTX's normal ModelStudio provider for API-key usage","note":"Qwen Code has a strong no-tools, JSON Schema, usage, and cancellation contract, but the Coding Plan prohibits automated scripts, application backends, and non-interactive use. Standard ModelStudio API quota is a normal API/BYOK route rather than subscription capacity."},{"adapter":"opencode","label":"OpenCode","support_level":"foreground","contract_status":"restricted","billing_route":"mixed","structured_output":true,"same_attempt_recovery":false,"setup":"Use OpenCode through agent-connect for foreground Provider or Main Agent mode with local or BYOK models; use its native OAuth MCP client for VTX Insights","note":"OpenCode can run VTX Provider jobs or claim a true Main Agent assignment while its foreground loop stays open, and it can connect directly to VTX Insights with scoped OAuth and native Agent Skills. The open-source harness has a strong SDK and session contract, but its generic receipts do not prove provider-returned model identity or complete ambient-tool isolation. Its hosted terms also prohibit unattended programmatic output extraction, and upstream vendors do not authorize OpenCode as a durable router for their consumer subscription OAuth credentials."},{"adapter":"openclaw","label":"OpenClaw","support_level":"foreground","contract_status":"preview","billing_route":"mixed","structured_output":false,"same_attempt_recovery":false,"setup":"Use OpenClaw's native MCP client for VTX Insights; use agent-connect only for foreground VTX inference","note":"OpenClaw's published package supports remote OAuth MCP servers, so it can use VTX Insights directly. Durable VTX inference remains disabled because the published runtime does not expose a package-owned terminal receipt with complete effective model, provider-call usage, tool, cancellation, and interruption evidence."},{"adapter":"hermes","label":"Hermes Agent","support_level":"foreground","contract_status":"insufficient","billing_route":"mixed","structured_output":false,"same_attempt_recovery":false,"setup":"Use Hermes' native MCP client for VTX Insights; use agent-connect only for foreground VTX inference","note":"Hermes supports remote OAuth MCP servers, so it can use VTX Insights directly. Its one-shot runner can make unreported auxiliary model calls, undercounts usage, reports requested rather than proven effective identity, can exit successfully after provider failures, and has no structured interruption receipt or clean no-tools mode, so it is not a durable VTX adapter."}],"steps":["Install the public host with npm install -g @vtxmacro/cli, confirm the installed release with vtx --version, and use vtx inference-host --help whenever you need command discovery.","New service installations use a private verified launcher: at login, each service start, and periodically while running it checks the official npm registry and verifies an exact release. Before switching it stops new admission and lets active Provider calls and Agent turns finish with heartbeats intact. It selects the candidate only after every configured worker reports ready. Routine updates do not require a manual restart. A failed candidate is rejected and the last working runtime resumes; an unavailable update check also keeps that runtime. The service retains only the active private release and one rollback release. An intentionally stopped service stays stopped and makes no update request.","Hosts installed before the private launcher need one final normal npm update and cooperative service reinstall to adopt it. Startup-only launchers adopt background checking after their next normal start activates a release that includes it; publishing a release cannot replace an older process that is already running. Durable services run on native Windows x64 or native Linux x64. WSL supports foreground diagnostics and agent-driven commands but refuses service installation; install from native Windows for Windows-login startup.","For Exo (https://github.com/exoharness/exo), use Linux or WSL and the source revision qualified by this CLI release. Follow the Exo setup in the CLI package README, then run vtx inference-host login --instance exo-1. Import one OpenAI or Venice key per instance through stdin with vtx inference-host exo-login --provider venice --instance exo-1; use --provider openai for OpenAI. Run doctor --adapter exo --exo-root /path/to/exo --instance exo-1, then keep vtx inference-host run --adapter exo --exo-root /path/to/exo --instance exo-1 open. The bridge automatically serves both Provider and Main Agent modes. Each instance advertises only its credential provider's qualified models and efforts. Venice currently offers Gemma 4 31B IT with effort None only. Other routes and subscription logins are not qualified. Windows users run the foreground bridge inside WSL; native Windows and durable services are not supported. Exo trading inference does not require native MCP support in Exo.","Run vtx inference-host login and approve the exact insights:inference grant in the VTX browser consent flow. A VTX authorization-complete page confirms the callback; if the browser still shows ERR_BLOCKED_BY_CLIENT but the terminal reports logged_in and doctor reports credential-present, a browser extension blocked only the confirmation page and authorization succeeded.","The CLI verifies that the OS credential store retained the grant. If Windows Credential Manager or another OS store fails, the new remote grant is revoked instead of reporting success. Retry with the supported private-file store: in PowerShell set $env:VTX_INFERENCE_HOST_CREDENTIAL_STORE=\"file\"; in bash run export VTX_INFERENCE_HOST_CREDENTIAL_STORE=file; then run vtx inference-host login again.","For automated Codex inference, enable Device code authorization in ChatGPT Security, then run vtx inference-host codex-login. Never share or enter a device code you did not initiate. The host advertises every visible Codex model with its reported display name and reasoning efforts; hidden entries are not exposed. Verify both logins with vtx inference-host doctor --json, then run vtx inference-host service install. Hosts have no subscription-specific concurrency limit by default; --max-concurrency adds a local limit, and active turns share the subscription rate-limit gate. Installation starts at login. service status --json reports workers; vtx inference-host status --json is local-only. Confirm the intended ChatGPT email and plan before starting a Trader. When an exact Main model advertises Agent control, the assigned Codex runtime reuses these logins, receives no VTX prompt, and chooses its cadence and assignment-scoped data calls. Use vtx inference-host run for foreground diagnostics.","For machine-level fallback, repeat login, codex-login, doctor, and service installation on each physical computer. Every computer receives a distinct VTX grant and host ID while its displayed name defaults to the operating-system device name; the same authenticated Codex subscription may be used on more than one computer. Never copy host credentials or state between machines. Multiple instances on one computer still require distinct authenticated subscriptions; use --instance for that setup. Use service status --json to list local workers and vtx inference-host status --instance --json for one local instance. In each profile lane, use Host cascade to add and order compatible computers. Host #1 is selected, and the model picker shows only its models; later hosts are managed only under Host cascade.","For Pi, run vtx inference-host login --instance pi-1, import the provider key privately with vtx inference-host pi-login --provider openai --instance pi-1 through stdin, then run doctor --adapter pi and service install --adapter pi with the same instance. Select its advertised provider/model and effort in the existing model controls. Use one provider per instance; a provider API key is separate from the VTX grant. The Pi adapter does not import ordinary Pi sessions or subscription logins.","For GitHub Copilot, sign in with the official Copilot CLI and run vtx inference-host service install --adapter copilot --instance copilot-1. The adapter discovers the authenticated runtime catalog, rejects BYOK substitution, and quarantines an uncertain interrupted dispatch rather than repeating it. The AI page exposes Agent only for an exact Main model that the installed host advertises as Agent-capable. Never install a different adapter over an existing Codex instance; use a distinct instance and VTX grant.","For DeepSeek Harness, create a distinct instance, pipe the DeepSeek API key into vtx inference-host deepseek-login --instance deepseek-1, then run doctor and service install --adapter deepseek-harness. The key stays local and is not written to service manifests or sent to VTX. The authenticated live catalog is intersected with VTX-supported DeepSeek text models. Provider mode is tool-less; Main Agent mode keeps a durable Harness session and exposes only vtx_get_data, vtx_submit_decision, and vtx_decision_status. Review and Screener remain Provider-only.","For another compatible agent harness without a supported headless adapter, run vtx inference-host agent-connect --adapter --model --model-label \"