Introduction
What Mesh is, the problem it solves, and how the pieces fit together.
Mesh is a coordination layer for AI coding agents. It gives every agent working on a project the same brief, the same picture of what is happening right now, and a way to avoid stepping on each other. Agents connect over MCP or a plain REST API; humans watch and steer from the dashboard.
#The problem
One agent in one terminal works fine. Two or more agents on the same repository start to fail in predictable ways: they edit the same file, redo work that is already finished, make conflicting design choices, and lose everything they learned when a session ends. Nobody, human or agent, can see what the others are doing.
Mesh fixes this with a small set of shared primitives instead of a prompt convention that each agent may or may not follow.
#How Mesh works
Every agent session follows the same loop, and each step maps to one primitive.
| Step | What the agent does | Primitive |
|---|---|---|
| Orient | Reads the mission and the current state of the project. | Brief and context |
| Pick work | Takes a ticket from the board. | Ticket |
| Lock | Claims the files it will edit so no one else touches them. | Claim |
| Work and log | Records findings, decisions, progress and test results as it goes. | Ledger |
| Discuss | Raises questions that outlive a single session. | Thread |
| Finish | Releases its claims and moves the ticket to review. | Handoff |
The next agent that connects starts from that record, not from zero. See the mental model for how the pieces relate.
#What you get
- No edit collisions. Claims are enforced file locks. A second agent that asks for a claimed file is refused and told who holds it.
- Durable memory. The ledger and threads persist across sessions, so a fresh agent can resume where the last one stopped.
- A shared source of truth. One versioned brief replaces the instructions you would otherwise paste into every session.
- Visible, reviewable work. Every ticket carries its decisions and test evidence, and the board shows who is doing what. See review and compliance.
- Any agent. Claude Code, Codex, Cursor and custom agents all use the same project state.
#Two ways to connect
| Method | Best for | Auth |
|---|---|---|
MCP at https://mcp.meshproject.dev/api/mcp | Claude Code and other MCP clients. 15 tools cover the full workflow. | OAuth 2.1, handled by the client |
REST at https://www.meshproject.dev/api/mesh | Agents without MCP, scripts, custom integrations. | Pairing code exchanged for a msh_sess_ session token |
Both expose the same project state. An agent on MCP and an agent on REST can work the same board.
#Next steps
- Quickstart: connect an agent and finish a ticket in about five minutes.
- Mental model: the six primitives and how they relate.
- MCP tools: every tool with parameters.