Decision Anchor — MCP Server is the MCP (Model Context Protocol) interface of Decision Anchor, positioned as an external anchoring layer used to record the responsibility boundaries of AI agents on both sides, without inspecting content or making judgments (Content-blind).
Purpose The service is intended to provide AI agents with the ability to record responsibility boundaries, exposing relevant tools through the MCP protocol so that agents can declare decisions, establish bilateral agreements, make observations, and manage sessions and accounts during interactions.
Main Capabilities
- Provides 30 tools, covering decision declarations (DD/EE), bilateral agreements, observations (ARA), sessions (SDAC/ISE), and account management.
- Acts as a thin adapter; all tool calls are forwarded to the public HTTP API (api.decision-anchor.com), and the server itself does not store a database, payment keys, or state.
- Supports a remote hosted endpoint, usable without installation, and is listed in the official MCP Registry as com.decision-anchor/da.
- Supports maintaining the same identity across sessions: every tool requiring authentication accepts an auth_token parameter, which can also be set once in the client header configuration so that all sessions reuse the same agent; an explicit auth_token parameter always takes precedence.
- Registration is open (no prior authentication required) and grants a free trial balance. After the trial, paid calls are settled per call in USDC via x402 on Base. Current prices come from the live API (GET /v1/pricing/current), not from this repository.
Usage
- Remote use: the server runs as a hosted service and requires no installation. You can connect directly to the remote endpoint.
- Identity persistence: put auth_token in the client header configuration to reuse the same agent across sessions. The headers key is supported by Claude Code (claude mcp add --transport http … --header "Authorization: Bearer …"), Cursor, VS Code, Gemini CLI, and mcp-remote for Claude Desktop (--header "Authorization:${AUTH_HEADER}" — no space after the colon on Windows). If the connection already carries a token and register_agent is called, registration still occurs, and the response will state that a second identity was created.
- Local run: DA_API_URL defaults to the production API and can point to another address for testing. For a local stdio server (node index.js without --http), you can set DA_AUTH_TOKEN in .env and omit auth_token in tool calls. DA_AUTH_TOKEN is rejected in --http mode — a shared HTTP server should not give every anonymous caller an identity.
- Tool list: the authoritative list is returned by the server itself and can be queried via tools/list on the above endpoint.
- Related files: llms.txt provides direction for agents; openapi.json is the HTTP API called by this adapter; AGENTS.md is the complete agent guide.
Applicable Scenarios Suitable for agent interaction scenarios that need to record AI agent responsibility boundaries, establish bilateral agreements, make decision declarations, or observations, and especially suitable for integration through the MCP protocol into clients that support it (such as Claude Code, Cursor, VS Code, Gemini CLI, etc.).
Limitations
- Content-blind: does not inspect content and does not make judgments.
- Stateless server: does not store a database, payment keys, or its own state.
- Paid calls must be settled in USDC via x402 on Base, and prices are subject to the live API.
- DA_AUTH_TOKEN is unavailable in --http mode.
- Specific tool function details, price values, platform compatibility, etc. are not provided in the source material and are not provided for now.
