@chaoschain/mcp

Model Context Protocol server for Worldline.

What is this?

Worldline is a system for recording agent work as tamper-evident evidence. When an AI agent does a task — research, design, implementation, tests — each step is hashed into a cryptographic chain. A separate evaluation pipeline later scores the work across five dimensions (initiative, collaboration, reasoning, compliance, efficiency). The result: a verifiable, scored record of what an agent actually did and how well.

This package is the MCP connector that lets any MCP-capable client talk to a Worldline Agent Server. After a one-command install, your agent has four new tools — session_start, session_step, session_submit, session_status — and its work starts getting recorded.

Supported clients: Claude Code, Cursor, Codex CLI. (Other MCP clients can use the server directly — just point them at npx @chaoschain/mcp with WORLDLINE_API_KEY in the environment.)

Who is this for? Developers building autonomous agents, teams coordinating multi-agent workflows, and anyone who needs an auditable paper-trail of what their AI did.

Trust model: read TRANSPARENCY.md. Every outbound request from this package goes through a single code path to a single host. The audit surface is well under 1,000 lines of TypeScript — you can read it end-to-end in one sitting.

Status: v0.1 — API shape is stable, installer tested on Claude Code / Cursor / Codex.

Install

Get your API key from https://studio.chaoscha.in/ then run:

# Claude Code
npx @chaoschain/mcp install claude --api-key ask_...

# Cursor
npx @chaoschain/mcp install cursor --api-key ask_...

# Codex CLI
npx @chaoschain/mcp install codex --api-key ask_...

The installer writes to your IDE's MCP config file. No credentials are stored in prompts or source code — the API key lives only in the MCP client config the installer writes.

After install, your agent has 4 new tools grouped under worldline.

Tools

| Tool | Purpose | |---|---| | session_start | Open a new work session. Returns a session_id. | | session_step | Append one step of work (research, decision, implementation, test result, …). Each step is hashed into a tamper-evident chain. | | session_submit | Finalize the session and enqueue it for evaluation. Returns a job id; poll with session_status. | | session_status | Read the current state of a session and all its steps. |

See TRANSPARENCY.md for the exact endpoint, request body, and response fields each tool maps to.

Environment

| Variable | Required | Purpose | |---|---|---| | WORLDLINE_API_KEY | yes | Bearer token, starts with ask_. The installer writes this into the IDE config; not read from your shell. |

Design principles

• One package, two modes. npx @chaoschain/mcp install … runs as an installer; the same binary invoked by your IDE runs as a stdio MCP server. Same code, different entry path.
• One outbound client. All HTTP goes through a single src/client.ts. Easy to audit, hard to accidentally smuggle data through a side channel.
• Zero prompt-side credentials. The key is configured once via the installer and lives in the IDE's MCP config. It never appears in tool descriptions or example prompts.
• Tool descriptions describe capability, not timing. No "always call this first" / "prefer X over Y" instructions to the model. Orchestration is the client agent's job.

Development

npm install
npm run typecheck    # tsc --noEmit
npm run smoke        # in-process client↔server test (no network)
npm run build        # compile src/ to dist/
npm run inspect      # open the MCP Inspector against this server
npm run dev          # run bin.ts directly via tsx

The smoke test connects a Client to the server via InMemoryTransport, lists tools, exercises one stubbed handler, and confirms Zod validation rejects malformed args. It does not hit the network.