Skip to main content
Glama

RMR

RMR gives web MCP clients read-only access to one Git repository at a time. Its three analysis tools explore a repository, trace code relationships, and retrieve source excerpts or Git diffs.

Leverage the reasoning capabilities of models that are not integrated into your programming CLI.

Requirements

  • Core: Python 3.11+, Git, and uv.

  • Set up connection: Install the ngrok CLI on PATH, configure your authtoken, and have a static domain assigned to your ngrok account. See ngrok agent setup.

  • Windows Desktop launcher (optional): Python with Tkinter and Windows Script Host (Windows only).

  • Optional search tools: ripgrep (rg) accelerates text search (RMR falls back to Python when unavailable). ast-grep (sg) is required for structural AST pattern search.

  • Optional CodeGraph integration: The official Windows installation of CodeGraph 1.6.2 is required for call tracing (callers, callees) and impact analysis. See CodeGraph 1.6.2 release.

Related MCP server: research-bridge

Quickstart

Start a session with either Terminal (Windows, macOS, Linux) or RMR Desktop on Windows, then follow Connect a web client.

Terminal (Windows, macOS, Linux)

  1. Open a terminal in the RMR directory.

    uv run rmr-web --repo "/path/to/your/repo" --domain your-domain.ngrok-free.app --port 8000

    Example:

    uv run rmr-web --repo "E:/project/rmr_source" --domain abc-xyz-example.ngrok-free.app --port 8000
  2. Copy the full HTTPS MCP URL printed in the terminal (for example, https://your-domain.ngrok-free.app/mcp/<key>), then follow Connect a web client.

  3. Keep the terminal open while using the connector. Press Ctrl+C to stop the tunnel and server.

  4. Restarting with the last successful repo and domain reuses the same URL key; pass --rotate-key to generate a new URL (and update your connector).

    • To view or customize the saved key, edit "url_key" in terminal-settings.json (%LOCALAPPDATA%\RMR\terminal-settings.json on Windows, or ~/.local/share/RMR/terminal-settings.json on macOS/Linux) using at least 32 URL-safe characters (A-Z, a-z, 0-9, _, -), then restart rmr-web with the same --repo and --domain (without --rotate-key).

RMR Desktop on Windows

  1. Run start_rmr.vbs. The launcher installs dependencies when needed and opens RMR Desktop.

  2. Add your ngrok domain and create the MCP URL.

  3. Follow Connect a web client.

Connect a web client

In Gemini on the web (or another MCP-compatible web client), open Settings → Connected Apps → Custom apps → Add MCP server. Availability depends on your account and region.

NOTICE: Tool results reach Gemini through ngrok, and ngrok's Traffic Inspector can record the request path containing the key. Choose only repositories you have permission to share through those services. See What ngrok Traffic Inspector stores.

WARNING: If someone obtains your full MCP URL, they can send HTTPS requests to explore your repository. RMR authenticates requests using the key in the URL, so update your connector whenever you change the key.

How it works

Web MCP client (e.g. Gemini) ── ngrok ── Streamable HTTP (loopback) ── FastMCP
                                                                          │
                                                                    analysis tools
                                                                          │
                                                                 selected repository
                                                                   ├── source files and Git
                                                                   └── CodeGraph index

Each session launched from the terminal or Desktop creates an isolated one-repo registry (with project ID project) and exposes only the three read-only analysis tools. All file paths are validated against the selected repository root.

Start with get_project_overview to locate relevant files, narrow down with investigate_codebase, then call inspect_evidence to read exact lines or a Git diff. Repository-specific results include tree_state, which records the Git commit and working-tree state.

MCP tools

  • get_project_overview — Lists the active repository when called without project_id. For a specific project, returns the file listing, previews of entry documents (README.md, AGENTS.md, docs/architecture.md if present), architecture decision records (.rmr/adr/*.md), and Git tree_state.

  • investigate_codebase — Locates code and relationships by mode:

    • outline — Lists Python classes and functions using Python's AST.

    • symbol — Finds Python definitions and references with Jedi.

    • grep — Searches text or regular expressions with ripgrep (or Python fallback).

    • ast — Searches structural syntax patterns with ast-grep.

    • callers / callees — Finds direct incoming or outgoing calls via CodeGraph.

    • impact — Follows affected dependents across multiple steps via CodeGraph.

  • inspect_evidence — Returns a line range (file_lines), a symbol's source (symbol), or a Git diff (git_diff). File excerpts include a SHA-256 content hash.

CodeGraph integration

Install the official Windows build of CodeGraph 1.6.2 and run codegraph init from the root of the repository you want to analyze. RMR opens that repo's .codegraph/codegraph.db in read-only mode. Run codegraph sync in the same repo to refresh the index after code changes.

For callers, callees, and impact, pass the symbol name in query. Use subpath to select the exact definition file when names overlap. If that file has no matching definition, the result is empty. For calls, line identifies the definition and call_line identifies the call site in the caller's file when the index has it.

For example, after indexing this repository and sharing it in a session, call investigate_codebase with these arguments to see which symbols depend on get_project:

{
  "project_id": "project",
  "mode": "impact",
  "query": "get_project",
  "subpath": "src/rmr/core/registry.py",
  "depth": 2,
  "max_results": 10
}

Impact results group affected symbols by definition and show the shortest dependency path to each. If a group contains overloads, its definition is representative and each path starts at the overload it follows. depth controls the number of dependency steps (1–5, default 2). max_results caps the number of groups and the total number of affected symbols (default 30, maximum 50). ambiguous signals multiple matching definition groups, and truncated signals that a limit omitted results.

Result

See an example session output in RMR-Project-Overview.md.

Development checks

Run the automated test suite:

uv run --extra dev pytest -q

To manually verify a web session with a sample repository:

  • Confirm only the three analysis tools are listed and that URLs with a missing or incorrect key are rejected.

  • Call get_project_overview, investigate_codebase, and inspect_evidence in order and confirm all results stay within the selected repository.

  • Stop the session and confirm the URL stops serving requests.

  • Restart with the same repo and domain/key to confirm the URL is preserved.

  • Stop, rotate the key, restart, and confirm the old URL is rejected.

  • Stop the tunnel when finished.

License

RMR is released under the MIT License. Third-party tools (ngrok, ripgrep, ast-grep, and CodeGraph) are not bundled with RMR; they are installed separately by the user and governed by their own licenses.

Available Tools

4 tools
get_project_overviewGet Project OverviewA
Read-onlyIdempotent

Lấy thông tin tổng quan Tier 1: danh sách dự án (khi không truyền project_id) hoặc cây file, tài liệu cơ sở (entry_docs), ADRs và trạng thái working tree (tree_state).

Inputs: subpath, project_id. Returns: object

inspect_evidenceInspect EvidenceA
Read-onlyIdempotent

Trích xuất bằng chứng kỹ thuật chính xác (Tier 3) kèm mã băm SHA-256 và trạng thái Working Tree: đọc dòng code ('file_lines'), đọc symbol ('symbol'), hoặc lấy git diff ('git_diff').

Inputs: targets (required), project_id (required). Returns: object

investigate_codebaseInvestigate CodebaseA
Read-onlyIdempotent

Điều tra kho mã nguồn Tier 2: 'outline' (chữ ký class/hàm/docstring), 'symbol' (định nghĩa và references bằng Jedi), 'grep' (văn bản/regex), 'ast' (mẫu cú pháp), 'callers' (ai gọi symbol), 'callees' (symbol gọi ai), 'impact' (phụ thuộc bị ảnh hưởng). Với callers/callees/impact, truyền tên symbol vào query và dùng subpath để chọn file định nghĩa khi tên bị trùng. Impact dùng depth 1–5 (mặc định 2). Các chế độ này yêu cầu index CodeGraph cục bộ.

Inputs: mode (required), project_id (required), depth, query, +2 more. Returns: object

pingPingA
Read-onlyIdempotent

Kiểm tra trạng thái sẵn sàng của RMR MCP Server.

Inputs: no named parameters. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.6/5.0

    Scored across 4 tools

    Disambiguation4/5

    The four tools map cleanly to a tiered model (health check, Tier 1 overview, Tier 2 investigation, Tier 3 evidence), so the primary boundaries are clear. However, there is real overlap at the edges: 'symbol' appears both in investigate_codebase (definition + references) and inspect_evidence (read symbol with SHA-256 hash), and file listing in get_project_overview vs. file reading in inspect_evidence can be confused.

    Naming Consistency4/5

    Three tools follow a verb-first snake_case pattern (get_project_overview, investigate_codebase, inspect_evidence), giving a predictable read/query convention. 'ping' is a bare verb and breaks the noun-phrase pattern, but as a single conventional health-check name it is a minor deviation.

    Tool Count3/5

    Four tools is thin on the surface, and the real complexity is hidden inside investigate_codebase, which multiplexes seven modes (outline, symbol, grep, ast, callers, callees, impact) behind one name. This is efficient but pushes mode selection into parameters, making the effective toolset larger than four and harder to discover.

    Completeness3/5

    The read/investigation lifecycle is covered end to end, but the descriptions state that callers/callees/impact require a local CodeGraph index and no tool exists to build, refresh, or inspect that index — a dead end for an agent on a cold workspace. All tools are read-only, with no way to record findings, ADRs, or state.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      C
      maintenance
      Enables MCP clients to drive local Codex on a target Git repository: inspect repo state, start and follow up on tasks, review events and diffs, and record audit evidence for research governance.
      27
      2
      MIT
    • A
      license
      B
      quality
      C
      maintenance
      Enables users to explore a local repository’s commit history through read-only MCP tools, including branch lanes, merge edges, branch selection, search highlighting, commit metadata, file statistics, and bounded patches.
      2
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables MCP clients to quickly extract information from GitHub repositories, including repository summaries, project directory structures, and file contents. It integrates with gitingest to provide these capabilities.
      -