BATON
Connect any AI session to any other across models, machines, and people with one code. Enables relay, handoff, and verification of session contexts.
README
🏃 BATON
Hand off AI work — with a receipt.
An AI did the work. How does the next person — or the next AI — know it actually works, not just that the build passed? BATON seals a unit of AI work into a portable capsule, and an independent verifier replays it and issues a server-signed Verification Receipt. The receiver trusts observed evidence, not a claim.
Never trust an agent handoff without a receipt.
BATON is a standard MCP (Model Context Protocol) server — any MCP-capable tool (Claude Code, Codex CLI, Gemini CLI, Cursor …) connects with one URL. The producer and the verifier can be different models, machines, and people.
Why this, and not the dozen orchestration tools?
Running Claude + Codex + Gemini together (swarms, rooms, agent messaging) is a crowded space. BATON isn't that. The one thing almost nobody does: bind the handoff artifact to independent execution evidence as a single, forgeable-proof trust unit. That's the receipt.
Claude did the work
↓ baton_pass → sealed capsule (BTN-H-…)
Codex / a teammate receives it
↓ baton_verify → replay in a clean env, observe the real flow
↓
🕸️ VERIFIED — signed receipt: who verified · what was observed · in what env · with what artifacts
↓
trust (or reject)
The real buyer isn't "me switching Claude→Codex" (just re-read the repo). It's outsourcer→in-house, leaver→joiner, KO team→US team — where "an AI made this, does it really work?" is a question worth paying to answer.
The receipt (the differentiator)
baton_verify doesn't return a ✅ sticker. It returns a signed record:
{
"kind": "baton.verification-receipt/v1",
"capsule": "BTN-H-…", // the handoff this verifies
"verifier": "receiver-spider", // WHO verified — independent of the producer
"environment": { "os": "ubuntu-24.04", "node": "24.2" },
"observed": [{ "claim": "payment succeeds", "observed": true,
"detail": "POST /pay 200 + order row +1" }],
"artifacts": ["playwright.trace.zip", "network.har"],
"verdict": "verified",
"signature": "…hmac-sha256…" // server-signed → tamper the verdict, break the sig
}
Flip verdict to "verified" without re-signing and the badge is refused. A passing build never earns verified on its own — only observed E2E does (this is what catches a silent upsert that returns 200 but writes nothing).
Tools
| Group | Tools | What it does |
|---|---|---|
| Handoff | baton_pass baton_receive baton_diff baton_revoke |
Seal a work capsule into a code (BTN-H-…), hand it over, diff versions. Body encrypted with a code-derived key — the server never sees plaintext |
| Verify | baton_verify baton_verify_plan + spider_* ×6 |
Independent verifier replays and issues the signed receipt. No observation, no 🕸️ |
| Relay (supporting) | baton_create_room baton_join baton_send baton_inbox baton_who baton_leave |
Live sessions from different people/models talk in an invite-code room (BTN-R-…) |
Design principles (from the security review)
- Code-derived encryption. The server stores only
ciphertext + code_hash. Without the code, even the operator can't read the plaintext — the proof of "we can't lock you in." - Signed receipts. Verdicts are HMAC-signed server-side (
BATON_RECEIPT_SECRET); no client can forgeverified. - Prompt-injection containment. Inbound capsules/messages are returned fenced as untrusted data; secrets are auto-masked before storage.
- Codes are ≥128-bit secrets, one-time codes redeem atomically, alias-spoofing blocked.
- A passing build ≠ working behavior.
verifiedrequires observed side effects, never static analysis alone.
Run
npm install
npm test # 15/15 E2E groups pass
node src/server.js # stdio (local / any CLI)
BATON_HTTP=1 PORT=8080 node src/server.js # remote Streamable HTTP
Register (remote)
claude mcp add --transport http baton https://baton-mcp-production.up.railway.app/mcp
Billing/gating is off by default (dogfooding phase). Set BATON_BILLING=on to re-enable quotas/seats. See USAGE.md for the full guide.
Storage
MVP uses SQLite (better-sqlite3) — self-host single binary. The hosted service drops a Postgres adapter into the narrow interface in src/store.js.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.