Claude Code
RepoMesh documents user-level MCP registration for Claude Code. The agent can call RepoMesh tools such as prepare_task and get_context before exploring unrelated files.
Multi-agent development · MCP integrations
One repository brain for Claude Code, Codex, Cursor and OpenCode
Switching AI coding agents should not mean starting repository discovery from scratch. RepoMesh is a shared local repository intelligence layer for MCP-compatible development tools.
Agent integrations, without changing how you code
Each agent continues doing the coding. RepoMesh answers repository-oriented context requests through its local MCP server.
OpenAI Codex
Configure Codex's MCP support to start RepoMesh from its local server command. Subsequent Codex sessions can ask for ranked repository context.
Cursor
RepoMesh documents Cursor MCP configuration and supporting guidance so the IDE agent can locate symbols, dependencies and tests with the same repository index.
OpenCode
Windows RepoMesh Desktop 0.0.49 includes the Connect OpenCode action and optional idle checkpoint hook. Install OpenCode Desktop, add a local repository in RepoMesh, connect OpenCode, then reopen it and check /mcps. The integration uses OpenCode V2's global mcp.servers.repomesh configuration. Existing MCP servers are preserved.
Other stdio MCP clients
A compatible client that can launch a local stdio MCP server may be configured manually. Compatibility depends on the particular client's MCP implementation.
Shared Task Handoff · Opt-in on Desktop
Keep repository state when switching agents
Available in Windows RepoMesh Desktop 0.0.49. The same local handoff store is accessible to connected MCP clients using the same RepoMesh home. RepoMesh doesn't control or run the coding agents.
Claude Code: capture on Stop
Install the optional Claude Code Stop hook with repomesh hook install claude. It records observed branch, HEAD and working-tree state after a turn; no transcript or model reasoning is captured.
Codex: opt-in Stop capture
In Windows Desktop 0.0.49, go to Settings → Shared Task Handoff → Automatic Stop hooks to install or remove the Codex Stop hook. PowerShell users can also run repomesh hook install codex. RepoMesh ignores assistant messages, transcripts and reasoning. Codex must trust the installed hook before running it.
Cursor: opt-in Stop capture
Cursor can read and explicitly save handoffs over MCP. Windows Desktop 0.0.49 also supports an optional Cursor Stop hook via repomesh hook install cursor. Its automatic record captures observed Git state, not the full task or chat transcript. Shared snapshot deduplication helps avoid duplicate events if Cursor also imports Claude hooks.
OpenCode: opt-in idle capture
After connecting OpenCode V2, you can separately install its optional session-idle checkpoint plugin with repomesh hook install opencode. It records bounded Git-state evidence when the client provides the session and repository directory; it does not capture reasoning or conversation history.
Check what changed
verify_task_state compares the saved Git snapshot with your current checkout. Matching Git state does not prove that tests passed or that a reported PR was merged.
Preserve decisions locally
Use save_project_decision and get_project_memory to retain rationale, track superseded decisions and surface potential conflicts across sessions.
You're in control.
Turn on Settings → Shared Task Handoff in Desktop. It is off by default, and turning it off blocks the handoff MCP tools without deleting history. Each client must use the same RepoMesh home. Hook installation is a separate explicit step for each supported client. This feature does not guarantee model token savings.
What RepoMesh does — and what it doesn't
Repository intelligence is shared; reasoning stays with the agent
RepoMesh indexes code, records useful repository knowledge, locates relevant source and tests, and helps tools avoid redundant exploration. It doesn't choose your model, generate application code on your behalf, or replace your coding environment.
How much context is returned?
RepoMesh attempts to rank and bound the information requested by a task. The scope depends on the index and query, so fewer model tokens are a goal, not a guaranteed outcome or invented metric.
What about ChatGPT?
Direct local ChatGPT integration is planned rather than presented as a shipped feature. Use one of the currently documented MCP client integrations or another suitable stdio MCP client.