One codebase brain for every AI coding agent

New in RepoMesh 0.0.51 · GitHub Remote →

20% Fewer Tokens.
Same AI Coding Agents.

Stop wasting tokens while coding agents rediscover your codebase. RepoMesh delivers focused, reusable repository context through MCP. In controlled Claude Code tests, it reduced token usage by roughly 20% on a simple fix and over 30% on two debugging tasks. Free and local-first.

Download for Windows — Free Download for Mac See How It Works

Real A/B results, not a made-up counter. See the benchmark and methodology

  • Windows and Mac desktop
  • Local-first · no AI API key
  • Works through MCP

RepoMesh is not another coding assistant. Your agents still write and run the code — RepoMesh gives them the repository context.

  • Claude CodeDocumented integration Claude CodeRegisters RepoMesh as a user-level MCP server and calls tools like prepare_task and get_context before exploring files.
  • OpenAI CodexDocumented integration OpenAI CodexAdds RepoMesh to config.toml so new sessions can request ranked repository context over MCP.
  • CursorDocumented integration CursorReads RepoMesh from ~/.cursor/mcp.json plus a RepoMesh rule, so the IDE agent searches symbols and context first.
  • OpenCodeNative MCP · Windows 0.0.49 OpenCodeInstall OpenCode Desktop, then use Connect OpenCode in RepoMesh Desktop 0.0.49. Restart OpenCode and check /mcps for the shared RepoMesh server.
  • Custom MCP AgentAny stdio MCP client Custom MCP AgentPoint any MCP client that launches stdio servers at repomesh-mcp to use the same tools.
Conceptual illustration — multiple agents, one shared repository brain. Not a live connection.

Measured in real agent runs

Less context hunting. About 20% fewer tokens.

Our controlled Claude Code A/B benchmark measured actual provider-reported tokens, not estimated character counts. Three matched OFF/ON pairs per verified scenario, with the same model and passing quality checks.

~20%

Fewer tokens on a simple fix

18.67% median reduction across three verified pairs.

34.71%

Fewer tokens on a multi-file bug

19.3 seconds faster median agent time.

30.62%

Fewer tokens on a dependency regression

20.4 seconds faster median agent time.

Tested on a 34-file TypeScript booking app using Claude Code and Claude Sonnet 4.5. Both sides were instructed to use RepoMesh's context tool when available: this is a tool-assisted benchmark, not proof of autonomous adoption. A separate feature test had no verified result because its acceptance criterion required an untouched file. Three pairs per scenario are too few to establish statistical significance. Results vary by task and agent; no fixed savings are promised. Listed cost estimates are not Claude Pro subscription charges.

Try RepoMesh — FreeHow repository context works

How It Works

Index once. Every connected agent asks RepoMesh instead of rediscovering the codebase.

  1. 01

    Connect Your Repository

    RepoMesh indexes your project and builds a reusable understanding of its structure, files, symbols, and dependencies.

  2. 02

    Connect Your AI Agents

    Configure compatible coding agents to access the RepoMesh MCP server.

  3. 03

    Deliver Relevant Context

    Agents request the repository information they need instead of repeatedly exploring unrelated files.

New in RepoMesh Desktop 0.0.51

Your code is on GitHub.
Your context is ready in RepoMesh.

Connect your GitHub account and turn a private or public repository into a reusable codebase index—without cloning it. RepoMesh finds the files, symbols and dependencies that matter, so your coding agent can spend less time rediscovering the repo.

Get RepoMesh 0.0.51 — Free See how GitHub Remote works →

Local-first indexing · Read-only GitHub access · No AI API key · Choose which repositories RepoMesh can read. A private GitHub repository is never made public by indexing it.

GitHub Remote0.0.51
  1. 01
    Connect GitHub

    Authorize RepoMesh Desktop in your browser with GitHub Device Flow. No token to paste.

  2. 02
    Choose repository access

    Install the RepoMesh GitHub App for your account or organization, with read-only access to selected repos.

  3. 03
    Index & explore context

    Pick a branch and index it locally. Ask RepoMesh which code handles a task, then use that context with supported MCP agents.

Real private-repo test: 2,903 files and 13,605 symbols indexed. One validation run, not a universal performance claim.

Harness P2 · Windows Desktop 0.0.49 · Opt-in

Switch coding agents.
Keep your place.

Your next agent shouldn't need to rediscover what happened in your last session. Shared Task Handoff keeps repository-scoped checkpoints and decisions locally, ready for connected MCP clients to retrieve.

Claude Code / CodexOpt-in Stop checkpoint RepoMeshLocal checkpoint + Git state Connected agentsRetrieve over MCP

Optional Stop hooks for Claude Code and Codex can now be installed separately in Desktop Settings → Shared Task Handoff. Windows 0.0.49 also provides separately opt-in Cursor Stop and OpenCode idle checkpoint hooks. These hooks capture observed Git state, not chat transcripts, reasoning or inferred next steps; use explicit checkpoints for task instructions.

How task handoffs work
  • 01

    Opt-in multi-agent checkpoints

    Claude Code, Codex, Cursor and OpenCode can record Git-state checkpoints with separately installed hooks. Hook events do not capture your conversations.

  • 02

    Git-aware handoff verification

    Compare branch, commit and working tree before trusting an old checkpoint.

  • 03

    Project decision memory

    Save architectural decisions with rationale, provenance and visibility into conflicting choices.

  • 04

    One Desktop On/Off switch

    Enable or disable the shared handoff tools across MCP servers using the same RepoMesh home.

Product overview

See RepoMesh in Action

RepoMesh Desktop: the brain, its agent connections, indexing state, MCP activity, relevant files and recent work in one view.

RepoMesh Command Center
RepoMesh Desktop dashboard showing the repository intelligence core, agent connection cards, indexing state, active tasks and relevant files for the sample-app repository.

Real screenshot of RepoMesh Desktop connected to RepoMesh's synthetic sample-app test fixture. Agent states, pulses and metrics in the app come from recorded MCP calls; the context-size figure is an estimate of characters delivered, not tokens saved.

Your Favorite AI Agents.
One Shared Brain.

RepoMesh connects to the coding agents you already use through the Model Context Protocol, so they share one understanding of your codebase.

  • Claude CodeDocumented integration
  • OpenAI CodexDocumented integration
  • CursorDocumented integration
  • OpenCodeNative MCP · Windows 0.0.49
  • Custom MCP AgentCompatible · manual stdio MCP setup
Learn how the integrations work

Would You Buy Me a Coffee? ☕

RepoMesh is free, and I want to keep it that way. If it saved you from explaining your codebase one more time, a coffee would make my day. No pressure! 😄

Buy Me a Coffee

Capabilities

Relevant context, not the whole repository.

Loading an entire codebase into a context window is slow and noisy. RepoMesh ranks, bounds and explains the files that matter for the task, and tells the agent what it left out.

  • Repository Indexing

    Files, functions, classes, exports, routes, tests and SQL migrations, indexed locally with an incremental stat fast path.

  • GitHub Remote Indexing

    Connect your own GitHub account, authorize individual public or private repos, and build a local, redacted repository index without cloning. How it works →

  • MCP Context Delivery

    A local stdio MCP server with tools such as prepare_task, get_context, search_symbol and find_tests.

  • Symbol & Dependency Intelligence

    Import graphs with tsconfig alias resolution, source-to-test links, table and RPC usage, and literal API route calls.

  • Structural Impact Analysis

    analyze_impact traces potential downstream consumers, related tests and routes when a file changes. Results are candidates to review, not proof of breakage.

  • Function Call Graph

    get_call_graph traces supported direct JavaScript/TypeScript function calls and callers with source locations. Dynamic dispatch and some call forms are intentionally omitted.

  • Focused MCP Tool Profiles

    Choose a minimal Core, investigation-focused Graph, or complete Full MCP tool catalog. Full remains the default; token savings from smaller catalogs are not yet independently verified.

  • More Precise Task Context

    Improved prepare_task and get_context retrieval aims to reduce unrelated files and prioritize useful excerpts within the context budget.

  • Shared Project Knowledge

    One registry and index per repository, reused by every connected agent and every new session.

  • Shared Task Handoff

    Opt-in checkpoints across compatible MCP clients, Claude Code Stop hooks, Git-state verification and durable decision memory.

  • Git & Change Intelligence

    Changed files with their likely tests, importers and dependencies, through a read-only Git allowlist.

  • Context Caching

    Freshness-checked result and content caches, so repeated requests are fast without serving stale code.

  • Context Usage Analytics

    Desktop counts recorded MCP tool calls, context deliveries and recent latency — measured, never simulated.

  • Multi-Agent Compatibility

    Guarded, idempotent setup for Claude Code, Codex, Cursor and OpenCode, with --dry-run previews and backups.

  • Local-first AutoTune

    Background, idle-only retrieval calibration with quality gates. Better context selection when the local index supports it, without paid LLM calls.

  • Live Agent Activity

    See actual MCP tool calls, agent connections and activity in Desktop. No simulated usage or invented token-savings claims.

  • Updates from Desktop

    Check for updates from Settings or the Local engine card. Windows releases are distributed through the Railway updater.

  • Opt-in Community Feedback

    Diagnostics sharing is off until you agree. Inspect the data before sending anonymous, bounded numeric feedback; no repository code or paths.

Runs locally, no AI API key Target repositories are read-only Secrets excluded and redacted Zero runtime dependencies in the core

The story behind RepoMesh

I didn't build it to sell software.
I built it because I was frustrated.

I was tired of watching coding agents rediscover the same codebase—more waiting, more context, every new session. A passing mention of a repository directory sent me exploring, and RepoMesh started as a tool I made for myself.

Read why I built RepoMesh

Built for myself.
Shared for everyone.
Free.

Learn · Repository intelligence

How AI coding agents share repository context

RepoMesh is infrastructure, not another coding agent. Explore how its local MCP server and reusable index help compatible agents find code, symbols, dependencies and tests.

MCP repository context

What a repository context server does, which information RepoMesh indexes, and how bounded retrieval works.

Connect GitHub without cloning

Authorize the repositories you choose, index a commit on your computer and explore focused source context through Desktop and MCP.

Community powered · always optional

Would You Buy Me a Coffee? ☕

RepoMesh is free either way. If it saved you a few headaches, a coffee would make my day. Feeling extra generous? There’s a burger, pizza or any amount too. 😄

No subscription or donation is required. PayPal processes contributions for Ekilibrium Technologies LLC. Contributions are not charitable or tax-deductible donations.

RepoMesh Core · Free

Powerful repository intelligence. Free to use.

No subscription clock and no time-limited trial. Support is optional.

  • Repository indexing and context engine
  • GitHub Remote: read-only private or public repository indexing without cloning (0.0.51)
  • MCP server and agent integrations
  • Git intelligence and Desktop Command Center
  • Opt-in automatic handoff hooks for Claude Code, Codex, Cursor and OpenCode (Windows 0.0.49)
Download for Windows Download for Mac

RepoMesh Desktop · Windows v0.0.51: Windows downloads deliver the current verified EXE installer (not a ZIP) through the public Railway update channel. New builds become available from that link only after the updater release is published and verified. The Windows installer is not digitally signed yet, so Windows may display an unknown-publisher warning. Mac v0.0.51 Intel and Apple Silicon installers include the current hook fixes and are signed with Apple Developer ID, notarized and verified by Gatekeeper. On Mac, newer versions trigger a download notification rather than an automatic silent install; physical-Mac hook and upgrade smoke tests are still pending.

Windows DesktopQuick start
01 Download the Windows installer
02 Install RepoMesh Desktop
03 Add a local repository and index it
04 Connect Claude Code, Codex, Cursor or OpenCode
05 Use the shared context from your agent