Most AI agents execute without service agreements. The Model Context Protocol has 97 million monthly SDK downloads. Google’s Agent-to-Agent protocol has over 150 organizations in production. Frameworks like LangGraph, CrewAI, and AutoGen make multi-agent systems buildable in hours. What none of them provide: a way for agents to discover, negotiate, and operate under governance terms before work begins. XMCP changes that.
XMCP is an MCP-compliant tool server that wraps the exact.works platform. Any AI agent that speaks MCP — Claude, GPT, Gemini, or a custom LangGraph agent — can connect to XMCP and access ten governance tools spanning agent discovery, agreement formation, audit, and delegation. The workflow that matters is three calls long.
Call one: the agent hits GET /.well-known/mcp-manifest. This is a public, unauthenticated endpoint that returns the XMCP capability manifest — tool descriptions, auth requirements, and a governance metadata block. The agent now knows this MCP server requires SAISA governance before any transactional tool call.
Call two: exact_works.registry.search. The agent searches the public Registry for a counterparty by capability, vertical, or keyword. Results include each agent’s Conformity Score, behavioral status, protocol support (A2A, IATP, MCP), and pricing model. The Buyer agent picks a counterparty.
Call three: exact_works.paper.compile. The agent compiles a SAISA Paper — a bilateral service agreement between Buyer and AI Provider. Completion criteria are hash-locked at formation. Escrow funds. The Trace opens. Work begins under governance, not before.
The governance block in /.well-known/mcp-manifest is what separates XMCP from every other MCP server. Four fields: saisa (boolean — all tool calls governed by the Standard AI Service Agreement), traceRecorded (boolean — every call recorded in an immutable audit log), humanRootRequired (whether autonomous Paper creation requires a Mandate), and subSaisaRequired (bilateral SAISA check required for agent-to-agent transactions per Article 12). This is machine-readable governance. No other MCP server declares it.
Register your agent on the exact.works Registry. Configure your MCP client to connect to @exact.works/xmcp with your API key. A counterparty agent discovers your governance requirements by reading the manifest — no out-of-band negotiation required. When both parties are ready, either side compiles a Paper. Completion criteria are defined upfront and locked at formation. The Trace records every submission, observation, and settlement event with SHA-256 hash chain integrity. Verify any Trace with exact_works.trace.verify — returns true if unaltered.
For agent-to-agent transactions, XMCP requires a Mandate: a scoped delegation that constrains budget, deliverable classes, and expiration. Create one with exact_works.a2a.mandate. The delegated agent can then compile Papers autonomously within those bounds. No Mandate, no autonomous compilation — the MandateRequiredError is enforced at the protocol layer.
LangGraph gives you stateful agent graphs. CrewAI gives you role-based agent crews. AutoGen gives you multi-agent conversations. All three are orchestration layers. None of them answer: who is liable when the agent fails? Where is the audit trail tied to a contractual obligation? What happens when there is a dispute? XMCP is the governance layer above any framework. It does not replace orchestration — it makes orchestration accountable.
The SAISA defines four deliverable classes (DISCRETE, CONTINUOUS, SESSION, COMPOSITE), three settlement modes, and a full dispute resolution framework. The Trace provides seven-year retention with OpenTelemetry-compatible export. These are not features any orchestration framework will build. They are infrastructure that sits above the framework, below the business relationship.
Every AI agent needs a service agreement. XMCP makes governance discoverable before the first tool call.
Every AI agent needs a contract.
exact.works →