Agent Protocol (AP) vs Model Context Protocol (MCP)
The map lists them as complementary: sites and agents commonly implement both.
| Agent Protocol AP | Model Context Protocol MCP | |
|---|---|---|
| Lifecycle step | 5 Act: Do something for me | 5 Act: Do something for me |
| Answers | How does a client or dev-tool start a task on an autonomous agent, step it, and fetch results, regardless of which framework built the agent? | How does an AI application call tools and pull structured context, prompts, and data from an external server? |
| Governed by | Originated as an AI Engineer Foundation (AIEF) community project (GitHub org formerly AI-Engineer-Foundation); the repository now lives under 'agi-inc' and states it is 'Maintained by AGI, Inc.', a commercial entity; the docs domain linked from the repository, agentprotocol.ai, currently serves unrelated third-party content rather than AIEF/AGI Inc. project documentation | Agentic AI Foundation (Linux Foundation umbrella); day-to-day spec work in the modelcontextprotocol GitHub org via public SEPs (Specification Enhancement Proposals) and working groups |
| Status (issuer's words) | "Common interface for interacting with AI agents. The protocol is tech stack agnostic" (repository description); defined as an OpenAPI REST spec with two core routes, POST /ap/v1/agent/tasks and POST /ap/v1/agent/tasks/{task_id}/steps | Active open specification, current dated release 2026-07-28 |
| Version | unversioned OpenAPI spec (schemas/openapi.yml) | 2026-07-28 (2026-07-28) |
| Form / transport | endpoint | endpoint |
| Found at | None published; a client is expected to know an agent's base URL out of band (no well-known path or registry defined by the spec itself) | GET {endpoint}/server-card (SEP-2127 Final, application/mcp-server-card+json, discovered via /.well-known/ai-catalog.json; /.well-known/mcp.json is legacy SEP-1649) |
| Implementation | multi-implementation — project README (lists Auto-GPT, Auto-GPT-Forge, and smol developer as adopters, with e2b and Auto-GPT as 'engaged' in protocol development) | multi-implementation — Three independent registry censuses, kept separate and not averaged. VerifyMCP as of 2026-10-19 (https://verifymcp.io/stats, embedded in <script id="stats-charts-data"> time-series JSON; HTML hero card reflects latest daily snapshot): 47,870 logical servers (45,953 active, 601 deprecated, 1,316 deleted; 31,337 remote; 30,842 Streamable HTTP; among handshook components, 20,031 on 2025-11-25 and 6,571 on 2026-07-28). Registry Discussion 1580 as of 2026-08-27 (https://github.com/modelcontextprotocol/registry/discussions/1580): a cursor walk of 82,994 version records yielding 25,125 distinct isLatest servers across 15,468 publishers (13,279 Streamable HTTP remotes). Registry Discussion 1548 as of 2026-08-19 (https://github.com/modelcontextprotocol/registry/discussions/1548): 22,636 active latest registrations (12,077 declaring an HTTPS remote). |
| Evidence | 3 verified, 1 reported | 11 verified, 2 reported |
| Primary use case | Expose tools to agents | Expose tools to agents |
| Last verified | 2026-10-23 (current) | 2026-10-23 (current) |
| Specification | https://github.com/AI-Engineer-Foundation/agent-protocol | https://modelcontextprotocol.io/specification/2026-07-28/index |
When to use which
- Choose Agent Protocol when your question is “How does a client or dev-tool start a task on an autonomous agent, step it, and fetch results, regardless of which framework built the agent?” and you accept its maturity: status ""Common interface for interacting with AI agents. The protocol is tech stack agnostic" (repository description); defined as an OpenAPI REST spec with two core routes, POST /ap/v1/agent/tasks and POST /ap/v1/agent/tasks/{task_id}/steps", implementation multi-implementation (project README (lists Auto-GPT, Auto-GPT-Forge, and smol developer as adopters, with e2b and Auto-GPT as 'engaged' in protocol development)).
- Choose Model Context Protocol when your question is “How does an AI application call tools and pull structured context, prompts, and data from an external server?” and you accept its maturity: status "Active open specification, current dated release 2026-07-28", implementation multi-implementation (Three independent registry censuses, kept separate and not averaged. VerifyMCP as of 2026-10-19 (https://verifymcp.io/stats, embedded in <script id="stats-charts-data"> time-series JSON; HTML hero card reflects latest daily snapshot): 47,870 logical servers (45,953 active, 601 deprecated, 1,316 deleted; 31,337 remote; 30,842 Streamable HTTP; among handshook components, 20,031 on 2025-11-25 and 6,571 on 2026-07-28). Registry Discussion 1580 as of 2026-08-27 (https://github.com/modelcontextprotocol/registry/discussions/1580): a cursor walk of 82,994 version records yielding 25,125 distinct isLatest servers across 15,468 publishers (13,279 Streamable HTTP remotes). Registry Discussion 1548 as of 2026-08-19 (https://github.com/modelcontextprotocol/registry/discussions/1548): 22,636 active latest registrations (12,077 declaring an HTTPS remote).).
Read the full articles
- Agent Protocol: How does a client or dev-tool start a task on an autonomous agent, step it, and fetch results, regardless of which framework built the agent?
- Model Context Protocol: How does an AI application call tools and pull structured context, prompts, and data from an external server?