MCP vs ACP: What Each Protocol Connects
MCP and ACP are both open protocols for AI systems, but they solve different problems. MCP connects an AI app to external tools; ACP lets agents communicate with each other. Here is what each one does and where they fit together.
Updated 5 min read
MCP (Model Context Protocol) and ACP (Agent Communication Protocol) both appear in conversations about AI systems, and the abbreviations are easy to confuse. They solve different problems. MCP is a standard for connecting an AI app to an outside tool — a calendar, a codebase, a task board. ACP is a standard for agents to communicate with other agents in a multi-agent system. Tools like fenbs work with MCP: an AI app connects to fenbs, reads the shared work record, and updates it.
If you are building or evaluating AI workflows, knowing which layer you are looking at helps you ask the right questions. Are you deciding how an assistant calls an external service? That is the MCP layer. Are you deciding how two agents divide a job and pass results between them? That is the ACP layer.
What MCP does
MCP, the Model Context Protocol, defines how an AI client — an app like Claude Desktop, Cursor, or ChatGPT — calls a server that exposes tools. The client asks the server for a list of available tools, the model decides which to call, the client sends the call, and the server returns a result. The model then uses that result to continue its answer.
The pattern is similar to a web API, but designed for AI apps: the tool descriptions are written so a model can understand what each one does and what it requires, without a developer writing glue code for every combination. A server that exposes a "list tasks" tool and a "create task" tool can be used by any MCP client without the client being written specifically for that server.
MCP connections use a browser sign-in (OAuth), so the user approves access in a pop-up rather than pasting a token into a config file. Once connected, the assistant acts as that user, within whatever access that user has on the connected service.
What ACP does
ACP, the Agent Communication Protocol, addresses a different problem: how do two AI agents — each with its own model, tools, and memory — exchange tasks and results? In a multi-agent system, one agent might break a job into pieces and assign each piece to a specialist agent. ACP defines the messages those agents use: how to send a task, how to ask for a status update, how to return a result, and how to stream partial output while work is in progress.
While MCP is primarily about a client reaching outward to a tool, ACP is about peers reaching laterally to each other. An agent using ACP might also use MCP on its own: it receives a task through ACP, then calls external tools through MCP to complete it, then returns the result through ACP.
A summary comparison
| MCP | ACP | |
|---|---|---|
| Main question | How does an AI app call an external tool? | How do AI agents communicate with each other? |
| Typical parties | AI client and a tool server | Agent and another agent |
| Direction | Client calls outward to a server | Peers communicate laterally |
| Sign-in | OAuth browser sign-in or header token | Varies by implementation |
| Common use | Reading files, calling APIs, querying databases | Multi-agent task delegation and result passing |
The two protocols are complementary. A multi-agent system might use ACP to coordinate the agents and MCP to give each agent its tools. Most people working with today's AI assistants are using MCP — it is already supported by Claude Desktop, Cursor, ChatGPT, Codex CLI, and others. ACP is newer and more common in purpose-built multi-agent platforms.
Why the names get confused
"MCP" and "ACP" share letters, sound similar in fast conversation, and both show up in the same discussions about AI agents. There is also genuine overlap: MCP serves agents as well as simpler assistants, and some agent frameworks define their own coordination patterns that blur the line.
The clearest way to stay oriented is to ask what is being connected. An AI app calling a calendar tool is an MCP question. Two agents handing a task between them is an ACP question. If an agent is doing both at once — taking a task from another agent and calling tools to finish it — both protocols are in play at the same time.
Keep the shared work record any connected agent can read
fenbs connects to AI apps over MCP. Each connected assistant — whether it is a standalone chat, an agent in a larger workflow, or a developer tool — reads the same projects, tasks, Decisions and rules, and Where we left off. Start with a free account and connect the first assistant from the Integrations page.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Does fenbs use ACP?
fenbs is accessible over MCP, not ACP. It is a shared work record that supported AI apps read and update through the MCP connection. It does not coordinate agents or pass tasks between them — that is the job of an ACP-based system or a multi-agent framework. If you are using an agent platform that speaks ACP, your agents can still reach fenbs if the platform also supports MCP tool connections.
The distinction matters for teams evaluating architecture. fenbs gives any connected agent the shared context — rules, current tasks, where the work stands — without becoming the coordination layer. The agent decides what to do; fenbs holds the agreed record of what has been done and decided.
How to do this with fenbs
- Create a free account and add your project to fenbs.
- Go to Integrations and follow the setup guide for your AI app — Claude Desktop, Cursor, ChatGPT, Codex CLI, Gemini CLI, GitHub Copilot, or Devin Desktop.
- Connect the app with the browser sign-in flow. The app stores the access token; you do not paste anything into a config file.
- Ask the connected assistant to read the project context — Rules, Where we left off, and current tasks — before starting work.
- If you are building a multi-agent system, set up MCP tool connections for each agent that needs to read or update the shared record. ACP handles the coordination between agents; MCP handles each agent's access to fenbs.
Reading more
For a broader look at how AI apps connect to tools, the MCP apps guide covers the clients that support MCP and what you can connect them to. Each vendor's documentation is the primary source for their current protocol version and server requirements.