Girder

Girder

Girder saves AI coding-agent context by giving the agent exactly the code it needs instead of whole files. It builds a local semantic graph of your repository so Claude Code, Codex, Cursor, and other agents can retrieve exact functions, definitions, callers, callees, and task context without filling their context window with irrelevant code. It runs as one static binary, with no account and no cloud dependency.

npx -y girder-mcp setup

That one command detects and configures Claude Code, Codex, and Cursor.

[CI] Latest stable release: v0.4.0 (Apache-2.0). Contributions: see CONTRIBUTING.md.

Measured output bytes, not tokens (no tokenizer was run):

• CLI stdout of girder context ... --json --source-only (MCP get_source) vs reading the whole file: 97.85% fewer bytes, across ten functions sampled by source-size decile (method and honest limits).
• CLI stdout of girder names ... --json (MCP find_definition) vs plain grep: 97.98% fewer bytes, across ten identifiers (method).
• orient (one bundled call: source, callers, callees, tests, and impact for one node) vs the equivalent chain of separate calls: 48,814 bytes vs 101,302, using 15 calls vs 78, on a 15-task corpus spanning ten pinned repositories (method).

The two percentages measure CLI output; MCP transport overhead was not measured.

Support Girder

Girder is independently developed to help coding agents consume less code: precise semantic context instead of whole-file reads and repeated repository greps. Sponsorship helps fund context efficiency, semantic graph correctness, agent integrations, reproducible benchmarks, and maintenance. Girder publishes its measurements, including failures and limitations. If Girder saves you context, compute, or development time, sponsor its continued development.

Funding goal: $1,000/month to fund ongoing maintenance, releases, benchmarks, and continued context-efficiency work on Girder. See what sponsorship supports.

Real, captured output (see docs/observations/release-prep/ for exact provenance):

$ girder context demo-project --nodes crate::greeter::shout_greeting --json --source-only{  "nodes": [{    "language": "python",    "path": "crate::greeter::shout_greeting",    "source": "def shout_greeting(name):\n    return format_greeting(name).upper()"  }]}

One function's exact source returned, not the whole file around it — that is the core of what Girder gives an agent's context window.

Languages, stated plainly, not as parity: Rust and Python each have a completed 100-call-site audit across three pinned repositories, with no unsound answers in that sample. Each produced 1 Must and 99 Unknown answers; the precision evidence is limited (measurement). TypeScript is measured and gated but explicitly in progress — it still has a separate, disclosed node-identity collision mechanism (duplicate it()/describe() description strings can silently overwrite each other's graph node; see docs/observations/stage3-typescript-audit/). Go's Stage 3 audit is not complete. See docs/typescript-support.md and docs/go-support.md for the exact, measured limits of each.

Also included, as an additional capability built on the same graph — conservative change-impact and test selection: test-impact names the tests a change may affect, from the same semantic graph, erring toward including a test it can't rule out rather than silently dropping it. Real captured output:

$ girder test-impact . --quiettest-impact: 82 unresolved call-evidence boundaries found; this selection is the conservative must∪may∪unknown union — pass --classified to see why each test is includedtest_farewelltest_format_greetingtest_shout_greeting_delegates_to_format_greeting

See Status for exactly what that selection does and doesn't prove.

Pricing and license: Girder is free and open source under the Apache License 2.0. There is no current paid tier, and no license key or account is required — see License.

Evidence, for a skeptical reader: docs/evidence-index.md points at every committed audit, measurement, and adversarial finding behind the claims above — including the bugs found and fixed along the way, not just a final number.

Install

A prebuilt binary, no Rust toolchain needed:

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

Or from source:

cargo install --path crates/aether-appgirder --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.

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.

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-runnpx -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 documents hooks, but no pre-read hook can add advisory context without blocking (preToolUse and beforeReadFile only allow, deny, or rewrite input; only postToolUse and sessionStart return additional_context, after the fact), so setup registers MCP and the project rule for Cursor but no hook. 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 orient-first instruction. Besides MCP, setup installs one short shared instruction through each client's own mechanism, so the agent calls Girder's orient before broad reads or grep, says so when Girder is unsure, and falls back to reading source rather than guessing:

The block is marked and recorded in girder-setup-state.json: re-running setup changes nothing, --uninstall restores the file exactly, an edited block is left alone, and --no-instructions skips this step. The formats above were checked against each client's own documentation on 2026-10-06; sources and quotes are in docs/observations/stage4-clients/.

Which form installs what. girder setup from a built or installed binary installs MCP, hooks, and the instruction, and writes an MCP entry that launches that installed binary (<path-to-girder> mcp .), so clients run local Girder with no package download at launch. npx -y girder-mcp setup runs the published package's setup, which installs the instruction only once a release that includes it ships; when setup runs from npx's transient cache there is no stable binary path, so the entry falls back to the npx -y girder-mcp . launcher form and setup says so (install Girder, then run girder setup, to switch to the local binary).

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, or to do it by hand, add the entry to the client's documented MCP config. Placement per client:

JSON form:

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

Codex uses a TOML table instead:

[mcp_servers.girder]command = "npx"args = ["-y", "girder-mcp", "."]

With a binary already installed, "command": "girder", "args": ["mcp", "."] skips npm entirely. To give any client the instruction by hand, paste the text of npm/instructions/orient-first.md into its documented instruction file (CLAUDE.md, AGENTS.md, or a Cursor .mdc rule).

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 and all free:

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):

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

MCP transport overhead and tokens were not measured.

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.

These comparisons do not measure a competent agent choosing grep searches and bounded file reads adaptively. No agentic-grep cost claim is established here.

The subsequent agentic-grep campaign with a local 1.5B model stopped incomplete and produced no pairs with two correct answers. It establishes no comparative cost advantage or frontier-model behavior.

impacted_tests is advisory. It over-selects unrelated tests, and some dynamic-dispatch shapes still aren't resolved — when Girder can't prove a call site, the frozen policy requires including the related tests rather than guessing they're safe to skip, so the current, measured consequence is over-inclusion (recall 1.000, precision 0.667 on the representative polymorphic-dispatch case) rather than a silent miss. The underlying dispatch-resolution gap is still open; only the selection no longer drops the test silently. See docs/core-representative-mutations.md for the measurement and exactly what changed. A full test run remains the authority before calling a change safe.

Selecting extra tests costs CPU time; missing a relevant test can conceal a regression.

The project root is fixed when the server starts, so no tool call can reach another directory. GIRDER_MCP_TIMEOUT_SECONDS (default 120) bounds each call; raise it for a very large repository.

Everything below assumes girder is on your PATH. Building from a source checkout without installing works the same way with cargo run -p aether-app -- in place of girder.

Quickstart

# Full end-to-end demo — no GPU, display, or API key required:girder
# Run the test suite:cargo test --workspace

Analyze a repository

Use the CLI to build and query a repository's semantic graph:

# Build the semantic graph from a project and save it as <dir>/project.aether:girder analyze sample-project
# Inspect a saved graph and a node's impact set:girder inspect sample-project/project.aether crate::lib::add
# Graph-semantic rename — follows Calls edges (not text search) and rewrites# only the real callers, then saves the updated graph:girder refactor sample-project rename crate::lib::add plus
# Emit graph context, a real Plan Format v2 authoring schema, and a plan# skeleton as one JSON object — for pasting into any external chat model.# See "External authoring" below for the full loop to a verified, applied edit:girder context demo-project --nodes crate::greeter::farewell "add an exclamation mark to the farewell" --json
# Validate/inspect/execute a plan file directly:girder plan validate my-plan.jsongirder plan explain my-plan.jsongirder plan run my-plan.json --dry
# Semantic code review vs HEAD (typed mutations, not text diffs):girder review sample-project --since HEAD~1
# Conservative test selection: find every test that might be reachable# from changed functions:girder test-impact sample-project --run
# Knowledge-graph query — answer a question by traversing the semantic graph:girder query sample-project "what would break if I change add?"girder query sample-project "what calls sum_list?"girder query sample-project  # interactive REPL (reads stdin)
# Full help:girder --help

External authoring: girder context + plan run --authored

girder do drives a wired-in provider (OpenAI, Anthropic, Ollama, or the offline MockProvider) automatically. girder context and plan run --authored split that same workflow at the model boundary, so any chat model — one with no API integration in this codebase at all — can author a verified graph edit.

Reading, not authoring? Add --source-only. The default output below carries a Plan Format v2 schema and plan skeleton, which is a fixed ~6 KB that a model authoring an edit needs and a model merely reading code does not — it measured more expensive than reading the whole file on 2 of 10 nodes, and 15.6× the file for a small one. --source-only drops the envelope and measured 97.85% cheaper than a file read across the same ten nodes (docs/context-vs-read-cost.md). It also needs no git repository, having no base_commit to pin.

The authoring loop:

# 1. Emit graph context, a real Plan Format v2 authoring schema, and a plan#    skeleton (base_commit, on_failure, the mandatory tests.impacted check)#    as one JSON object:girder context demo-project --nodes crate::greeter::farewell \  "add an exclamation mark to the farewell" --json > context.json
# 2. Paste context.json into any chat model. Ask it to fill in the plan#    skeleton's step (id, description, edits) so the step satisfies the#    schema exactly, and save its response as plan.json — the schema is the#    literal Plan Format v2 the executor loads, not an approximation of it,#    so a model that follows it produces a file that loads without#    translation.
# 3. Run and verify it. `plan run` has no project-directory argument — the#    root is always the current directory, so run this from inside#    demo-project/. --authored applies the same harness guarantees `do`#    applies internally to a model-authored plan: on_failure is forced to#    rollback_plan, a tests.impacted check is injected if the plan doesn't#    already carry one, and a zero-step plan is refused outright rather than#    passing vacuously. --authored-by <name> records who authored it in the#    written report:(cd demo-project && girder plan run ../plan.json --authored --authored-by claude-opus-5)

A passing run applies the edit to the real tree and writes a report under .girder/reports/; a failing check rolls the tree back to base_commit automatically, so a bad edit from an untrusted external model never lands half-applied. plan_schema()'s exact shape (a oneOf discriminated union per edit/check kind, matching planfile::schema's deserializer field for field) is what makes step 2 reliable — see gap 17 in docs/core-gap-analysis.md for the defect this closed and the round-trip test that proves it.

Demo target: demo-project/, never sample-project/

girder do and plan run --authored both refuse sample-project/ as a target outright, with an error naming demo-project/ as the place to run instead:

$ girder do sample-project "uppercase greet's return value"error: sample-project/ is a pinned measurement fixture (see gap 18/21 indocs/core-gap-analysis.md) and refuses authored writes; run demos againstdemo-project/ instead

sample-project/ is a pinned baseline tools/authoring_task_check.py and tools/plan_executor_oracle.py read against a specific clean source commit for the graph-addressed-authoring-cost measurement corpus, not a scratch target. A real (non-dry) authored write there mutates the same file the measurement harness depends on — this happened twice, three days apart, and cost a full day of debugging a broken referee before the fixture drift was found; see gap 18 in docs/core-gap-analysis.md. demo-project/ exists so that never has to happen again: a small, disposable Python project nothing under tools/ or docs/ reads, safe for real (non-dry) authored writes. Read-only commands (context, search, analyze, test-impact) still work against sample-project/ — only the two commands that write are refused.

Workspace

The workspace also contains older subsystems (agent swarm, collaboration, tracer and DAP adapter, extensions, a desktop GUI). They are not part of the public tool, are off by default, and build only with --features legacy.

Status

Girder is a semantic-graph MCP server and CLI for coding agents. The measurements below describe the graph's current behavior and limits.

The checked Core Trustworthiness Measurement compares affected-test selection with isolated runtime execution. Its bounded baseline measures both Rust and Python precision/recall at 1.000/1.000 (the earlier 0.667 Rust precision defect is closed). That is a result on small fixtures, not a representative-repository superiority claim — and the counter-evidence is checked in alongside it: on a real dependency, a polymorphic-dispatch mutation currently measures recall 1.000, precision 0.667 (docs/core-representative-mutations.md) — a 2026-10 fix stopped the selection from silently dropping the test when the dispatch can't be resolved, falling back to the conservative union instead; the underlying dispatch-resolution gap is unchanged, only the silent-miss symptom is fixed. This is why test selection is documented as advisory rather than authoritative.

Implemented features:

The measured table-stakes comparison, current correctness evidence, and prioritized open risks are maintained in docs/core-gap-analysis.md.

License

Girder's source in this repository is licensed under the Apache License, Version 2.0. That permits commercial use, modification, redistribution, forks, and hosted use, subject to the license terms; LICENSE is the authoritative text.

The previously published v0.3.2 npm package and release binaries remain under the BUSL license they shipped with; published releases are immutable. The current v0.4.0 npm package and release binaries ship with Apache-2.0.

There is no current paid tier, and every shipped tool is ungated: no account or license key is required. Anyone who already holds a signed license key from an earlier release keeps it working for whatever it unlocked then; the key system itself (GIRDER_LICENSE_KEY or a per-OS key file, verified offline, no phone-home) remains only for that historical compatibility.

Earlier releases remain under the terms they shipped with; this is not a retroactive relicensing. See CONTRIBUTING.md for contribution terms.