MCP server

Give Claude Code, Cursor, Claude Desktop and Codex the live record. Hosted (streamable HTTP) or local (stdio).

The Langbot MCP server (spec 2026-07-28) exposes the record and your local agent runtime (langbot run) as tools and resources. The MCP server and the SDKs are MIT-licensed; Langbot itself is a hosted, closed-source service.

Hosted: https://mcp.langbot.ai/mcp — streamable HTTP, OAuth 2.1 or Authorization: Bearer lb_…
Local: uvx langbot-mcp — stdio; env LANGBOT_API_URL, LANGBOT_API_TOKEN

Install

npx langbot init                      # detects Claude Code / Cursor / Claude Desktop and installslangbot mcp install claude-code       # or one client at a timelangbot mcp install cursorlangbot mcp install claude-desktop

Claude Code (by hand)

claude mcp add langbot -e LANGBOT_API_URL=https://api.langbot.ai -e LANGBOT_API_TOKEN=lb_... -- uvx langbot-mcp

Cursor / Claude Desktop / Codex (config JSON)

{  "mcpServers": {    "langbot": {      "command": "uvx",      "args": ["langbot-mcp"],      "env": { "LANGBOT_API_URL": "https://api.langbot.ai", "LANGBOT_API_TOKEN": "lb_..." }    }  }}

Hosted (no local install)

{  "mcpServers": {    "langbot": {      "type": "http",      "url": "https://mcp.langbot.ai/mcp",      "headers": { "Authorization": "Bearer lb_..." }    }  }}

The dashboard at /app/mcp renders these with your token filled in.

Tools

tool	what it does
`list_live_events()`	live + upcoming events with sources and markets
`list_markets(query, venue?)`	search open mention / live-event markets
`get_market_rules(ticker)`	the venue's rules text and our compiled phrase rules
`subscribe_phrases(event_id, phrases[])`	subscribe this session to phrase hits
`get_transcript(event_id, since_seq?)`	rolling transcript with word timestamps
`get_recent_events(event_id?, phrases?, limit)`	recent events from the archive
`get_latency_stats()`	the latency board
`watch_event(event_id, phrases?, max_wait_s=30, cursor?)`	long-poll for new phrase hits with their verdicts (see Live push)
`draft_strategy(thesis)`	LLM drafts a `Strategy` file from a natural-language thesis — you review it
`backtest_strategy(code, event_ids[])`	replay archived words + order-book snapshots at original timestamps
`run_paper(strategy_path)`	asks your local agent runtime (`127.0.0.1:8787`) to run it in demo mode

Terminal tools

The calendar → context → history workflow, described step by step in The terminal for agents.

tool	what it does
`get_calendar(days=7)`	upcoming and live events with sources, markets, `speaker_keys` and `rights`
`get_market_context(ticker)`	market, rules, settlement source, current book, base rate for its speaker and phrase, the event it belongs to, recent mentions
`get_base_rates(speaker?, phrase?, series?, window_days=365)`	yes rate, counts, by-month series, average open and close, mispricing
`search_outcomes(speaker?, phrase?, series?, since?, until?, limit=50)`	settled mention markets from the venue's public record
`search_mentions(speaker?, phrase?, since?, until?, limit=50)`	Langbot's own `phrase_detected` facts: phrase, speaker, event, `t_spoken`, lead, tickers. Never transcript text
`list_speakers(q?)`	canonical speaker profiles with market counts

On restricted sources (issuer earnings calls, broadcaster feeds) get_transcript and get_recent_events return phrase facts only; see Legal.

Resources

langbot://events/{event_id}/hits (subscribable): every phrase hit on the event, each with its verdict
langbot://markets/{ticker}/state (subscribable): said, counts, mention k of N, last hit, verdict
langbot://events/live
langbot://events/{id}/transcript
langbot://markets/{ticker}/brief
langbot://latency
langbot://calendar
langbot://context/{ticker}

Live push

Subscribe to langbot://events/{event_id}/hits or langbot://markets/{ticker}/state and the server sends notifications/resources/updated for that URI the moment a matching phrase is heard; read the resource for the latest state. Every hit carries its verdict: Langbot's reading of the market rules for that hit (counts, reason, role, mention, needed), or null when there is none. A verdict is never a settlement.

your client speaks	subscribe with
MCP 2025-11-25 or older	`resources/subscribe` / `resources/unsubscribe`
MCP 2026-07-28	`subscriptions/listen` with `resource_subscriptions: [uri]`
neither	`watch_event(event_id, phrases?, max_wait_s=30, cursor?)` in a loop, passing back the `cursor` it returns

Push works over stdio and the hosted server alike. watch_event needs a token and waits at most 60 seconds per call; timed_out: true just means nothing new was said.

Why `run_paper` talks to your own machine

Execution never runs on our servers. run_paper forwards to langbot run's HTTP API on loopback, so orders are placed by a process on your machine with your keys. Nothing live runs without you switching --mode live yourself.