Girder

Girder gives coding agents exactly the code they need, instead of whole files. It parses your repository into a living semantic graph — functions, definitions, call edges — and answers questions against that graph: exact function source, callers and callees, impact analysis, minimal test selection, and verified graph-addressed edits. It is one static Rust binary that any agent can drive over MCP, plus an optional native IDE.

npx -y girder-mcp setup --dry-run
npx -y girder-mcp setup

Languages: Rust, Python, TypeScript/TSX, and Go. Rust and Python are the most mature; TypeScript and Go are measured and gated, with their limits written down (TypeScript, Go).

Tiers: the free tier is permanent and needs no account — get_source, find_definition, search_code, ask_codebase, review_changes, and orient on a single repository. The impacted_tests tool needs a paid license. Keys are verified offline; the binary never phones home.

On the committed 15-task orient measurement, one bundled call per task returned 48,814 output bytes versus 101,302, using 15 calls versus 78 for the equivalent command chain, with 37/37 gated checks passing after the disclosed fixes. See the post-fix observation and limitations. These are output bytes, not tokens, and the baseline is Girder's separate commands.

Install

A prebuilt binary, no Rust toolchain needed:

curl -fsSL https://raw.githubusercontent.com/dhishwasher/Girder/main/install.sh | sh
girder --version

Or from source:

cargo install --path crates/aether-app
girder --help

This builds the default headless profile and installs the girder binary to ~/.cargo/bin (make sure it's on your PATH). No GPU, display, network, or API key is required — the default AI provider is an offline MockProvider. The GUI and live AI providers are opt-in Cargo features not included in a plain install; see The GUI and Local-first AI below.

Windows installer

Each Windows release also includes Girder-<version>-setup.exe. It installs for the current user under %LOCALAPPDATA%\Programs\Girder, adds Girder to the user PATH, creates a Start Menu shortcut, and does not request administrator access. Open a new terminal after installation so it sees the updated PATH. The installer includes the desktop GUI, and its Start Menu shortcut opens it. The archives and npm installation continue to provide the headless CLI.

The installer is not code-signed yet, so Windows SmartScreen will warn on first run. After downloading the installer from the GitHub release, double-click it, choose More info on the “Windows protected your PC” dialog, verify that the app is Girder and the publisher is shown as unknown, then choose Run anyway. If those details do not match, cancel instead.

Search accuracy

Exact-symbol lookup is Girder's strongest search path. Natural-language intent search is experimental: it reached 41.9% top-1 and 77.4% top-5 accuracy on the committed 31-item corpus, below the precommitted 75% and 90% thresholds. See the observation and policy.

Competitor benchmark

On the 20-task modest cross-language base set, Girder passed 5/20 versus Ripwire's 7/20 — a loss. But across the 60 attempts where both products passed, Girder returned 35,764 bytes versus Ripwire's 158,514 — 77.4% fewer — at identical call counts (120 each). GitNexus could not be scored: its install exceeded the frozen 1 GiB ceiling on this measurement machine, which is a fact about install footprint on a constrained host, not a claim about its quality.

No product in this benchmark reached full correctness, and the benchmark does not establish an overall best product. See the full report for methodology, per-language results, and everything the numbers above don't establish.

Use it from an AI coding agent

girder mcp serves the read-only graph commands over the Model Context Protocol, so an agent can ask about your codebase instead of reading files into its context window.

Preview the changes, then configure detected agents:

npx -y girder-mcp setup --dry-run
npx -y girder-mcp setup

setup detects Claude Code (~/.claude or an existing in-home project .mcp.json), Codex (~/.codex or an in-home CODEX_HOME), and Cursor (~/.cursor). It merges the girder MCP entry into each detected agent's documented config and installs nested PreToolUse and PostToolUse hooks for Claude Code and Codex. Pre hooks cover structured Read, read_file, and mcp__.*__read_file events; post hooks cover Claude's Edit|Write|NotebookEdit and Codex's apply_patch|Edit|Write edit names. Shell commands are intentionally not parsed. Cursor is MCP-only: its documented post-edit hook input and output semantics do not establish the standalone additional-context protocol used by this launcher, so setup does not register a hook there. A generic MCP client has no universal config path and is reported as not detected. Setup never writes outside your home directory. Existing foreign girder entries remain untouched, including with --force; girder setup --uninstall removes only setup-owned changes.

The packaged hook launcher forwards each event to the pinned native girder hook executable and fails open on errors. Only standalone PreToolUse JSON is forwarded on stdout; post-edit native diagnostics pass through stderr, and MCP continues to use JSON-RPC on its own stdout. The read advisory only gives guidance when project.aether already exists. The post-edit advisory uses that saved graph snapshot, does not scan or rebuild the project, and describes the last analyzed version of changed files. See the setup and client path guide for config paths, ownership, and the manual JSON fallback.

For clients setup cannot detect, add this entry to their documented MCP config manually:

{
  "mcpServers": {
    "girder": {
      "command": "npx",
      "args": ["-y", "girder-mcp", "."]
    }
  }
}

With a binary already installed, "command": "girder", "args": ["mcp", "."] skips npm entirely.

Girder 0.2.6 and later can opt into a cached graph generation with girder mcp . --watch or npx -y girder-mcp . --watch. The server keeps one validated graph generation in memory and incrementally reparses changed files. The committed watcher measurement matched fresh cold analysis after all 45 mutations while reusing 98.70% of file extractions; the claim is limited to preserving cold-analysis resolution while reusing parsing. See the watcher result and limitations.

Seven tools, all read-only:

What that saves, and what it doesn't

orient bundles what get_source + ask_codebase (callers, callees, and impact) + impacted_tests otherwise answer across 5-6 separate calls into one. On a 15-task corpus spanning ten pinned repositories, that one call used fewer aggregate bytes than the chain it replaces (48,814 vs 101,302, a 0.48 ratio) while cutting 78 round trips to 15 — one per task — and, after two disclosed defects were fixed, 37 of 37 gated checks pass. The first run found impacted_tests --quiet silently dropping non-Rust/Python test names (orient's own test-coverage section did not share the bug, which is how it was found); that filter is now removed. Its natural-language intent input still inherits search_code's accuracy — all three intent tasks in this corpus resolved to the wrong node, unchanged and out of scope for this fix — but orient's confidence heuristic, which originally caught none of the three, now flags all three "confidence": "low" with candidate scores attached, at the cost of also flagging some correct resolutions when a runner-up is close. See docs/orient-tool.md and the committed policy / original observation / post-fix observation.

Two additional precommitted measurements, both counting bytes of command output rather than tokens (no tokenizer was run):

• get_source against a naive whole-file-read baseline: 97.85% fewer bytes across ten functions sampled by source-size decile, cheaper on all ten (docs/context-vs-read-cost.md).
• find_definition against a plain-grep baseline: 97.98% fewer bytes across ten identifiers (docs/names-cost.md).

Both are single-repository measurements. The direction is structural — files are much larger than the functions in them, and grep returns every mention where find_definition returns only declarations — but the exact percentages are not portable.