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.

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 it misses tests reached only through dynamic dispatch (measured: recall 0.000 on a polymorphic-dispatch case, docs/core-representative-mutations.md). 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 (graph, builder, AI router, agent swarm, debugger):
cargo test --workspace

# Optional: compile live providers and run the real debugpy adapter test:
cargo check -p aether-ai --features live-providers
python3 -m pip install debugpy
cargo test -p aether-dap --test debugpy -- --ignored

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

# Concept search: rank functions by relevance to a natural-language query:
girder search sample-project "sum numbers in a list"

# 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

# Preview what the swarm would build — graph-aware Planner only, no code written:
girder swarm-plan sample-project "add user authentication"

# Dispatch the agent swarm on a project with a natural-language intent:
girder forge sample-project "add a subtract function"

# Author a single graph-addressed plan step via a wired-in model (the
# offline MockProvider by default) and execute it through the same
# verified plan executor as `plan run` below, repairing from check
# failures automatically. Never point this — or `plan run --authored` —
# at sample-project/: it is a pinned measurement fixture, not a demo
# target, and both refuse it outright (see "Demo target" below):
girder do demo-project "add an exclamation mark to the farewell"

# 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
# that isn't wired in as a provider. See "External authoring" below for
# the full loop from here 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. `--authored` is for a plan
# an external model wrote by hand (see "External authoring" below); without
# it, `run` executes a plan exactly as authored (used internally by `do`):
girder plan validate my-plan.json
girder plan explain my-plan.json
girder plan run my-plan.json --dry

# Semantic code review vs HEAD (typed mutations, not text diffs):
girder review sample-project --since HEAD~1

# Minimal test selection: find every test 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)

# Start a graph-native collaboration history, give another replica its own actor,
# record that replica's current source graph, and deterministically merge it:
girder collab init sample-project alice alice.aetherc
girder collab fork alice.aetherc bob bob.aetherc --approve
girder collab sync sample-project bob.aetherc
girder collab merge alice.aetherc bob.aetherc merged.aethercb
girder collab materialize merged.aethercb merged.aether
# Membership changes are causal operations and require explicit approval:
girder collab member add alice.aetherc carol --approve
girder collab member remove alice.aetherc carol --approve

# Or exchange deltas in a mutually authenticated live loopback session.
# Secret contents are generated with private permissions and never printed.
# Group-secret-only operation remains available as a migration mode:
girder collab secret collaboration.secret

# Strict identity mode additionally pins each roster actor to an Ed25519 key.
# Generate each actor's private/shareable-public pair, compare the printed
# SHA-256 fingerprints out of band, and approve the exact peer fingerprint:
girder collab identity generate alice.aetherc \
  alice.identity alice.identity.pub
girder collab identity generate bob.aetherc \
  bob.identity bob.identity.pub
girder collab identity show bob.identity.pub
girder collab identity trust alice.trust \
  bob.identity.pub --approve <bob-fingerprint>
girder collab identity trust bob.trust \
  alice.identity.pub --approve <alice-fingerprint>
# Sign existing local-authored history now (strict host/join also does this
# in memory before exchange), then audit a fully attested bundle offline:
girder collab identity attest alice.aetherc alice.identity
girder collab identity attest bob.aetherc bob.identity

girder collab host alice.aetherc 127.0.0.1:7331 \
  --identity-file alice.identity --trust-store alice.trust \
  --secret-file collaboration.secret --discovery-dir .bitcode/peers \
  --presence "reviewing parser changes"
girder collab discover bob.aetherc .bitcode/peers \
  --secret-file collaboration.secret
girder collab join-peer bob.aetherc alice .bitcode/peers \
  --identity-file bob.identity --trust-store bob.trust \
  --secret-file collaboration.secret --presence "running transport tests"
# An explicit address remains available when local discovery is not in use:
girder collab join bob.aetherc 127.0.0.1:7331 \
  --secret-file collaboration.secret \
  --identity-file bob.identity --trust-store bob.trust
# Verify every retained non-bootstrap operation against the current local pins:
girder collab identity verify \
  alice.aetherc alice.identity alice.trust
# To rotate your own key, first generate a new pair, then record a dual-signed
# causal transition while both private keys are available. Peers can then rotate
# their current pin before the next strict session:
girder collab identity generate alice.aetherc \
  alice-new.identity alice-new.identity.pub
girder collab identity rotate-local alice.aetherc \
  alice.identity alice-new.identity \
  --from <old-alice-fingerprint> --approve <new-alice-fingerprint>
# Peer trust rotation/removal also requires the exact reviewed fingerprints:
girder collab identity rotate alice.trust \
  bob-new.identity.pub --from <old-bob-fingerprint> --approve <new-bob-fingerprint>
girder collab identity remove alice.trust bob \
  --approve <new-bob-fingerprint>
# Successful sessions persist both peers' causal acknowledgements. Once every
# active member has acknowledged superseded history, prune it conservatively:
girder collab compact alice.aetherc
# Rebuild remote whole-file projections, show semantic/file changes and
# conflicts, then explicitly validate and journal-commit the reviewed bytes:
girder collab review sample-project alice.aetherc
girder collab apply sample-project alice.aetherc --approve

# Real Python execution tracer — records every variable at every line/call/return:
girder debug script.py
girder debug script.py --what-if x=10 at 2

# DAP adapter dry-run: resolve graph breakpoints without launching an adapter:
girder dap script.py --dry-run

# Generate and review an extension recipe without changing the project:
girder extension sample-project generate "show call impact"

# Grant the exact recipe digest/capabilities, then manage its lifecycle:
girder extension sample-project generate "show call impact" --approve
girder extension sample-project list
girder extension sample-project disable dev.bitcode.generated.show-call-impact
girder extension sample-project remove dev.bitcode.generated.show-call-impact

# Install a hand-authored declarative recipe after the same explicit review:
girder extension sample-project install recipe.json --approve

# Browse the built-in reviewed marketplace and inspect a listing:
girder extension sample-project marketplace search impact
girder extension sample-project marketplace show org.bitcode.impact-navigator

# Regenerate a reviewed intent for this project, preview its capability delta,
# then explicitly approve the adapted recipe:
girder extension sample-project marketplace adapt org.bitcode.impact-navigator
girder extension sample-project marketplace adapt org.bitcode.impact-navigator --approve

# Portable catalogs use the same bounded parser and print a catalog fingerprint:
girder extension sample-project marketplace list \
  --catalog marketplace/girder-extensions.json

# Full help:
girder --help

analyze/forge walk every .rs, .py, .ts, .tsx, .mts, .cts, and .go file (skipping target, .git, …), build the graph with directory-aware module paths, resolve free and receiver-qualified method calls across files, and persist the .aether graph. Supported languages are Rust, Python, TypeScript/TSX, and Go. The TypeScript and Go graph surfaces are measured and gated, but remain less mature than Rust and Python: TypeScript meets its precision and recall gates, while Go currently records one false negative (micro-recall 0.954545 against a 1.0 gate). See the honest support boundaries and results for TypeScript and Go, with the committed observations for TypeScript and Go. Rust module-scope imports retain renamed symbol identity across bounded public re-export chains, including crate-root and mod.rs facades, so collisions are resolved by exact path while ambiguous or cyclic aliases stay unlinked. Rust parameter annotations and direct type-qualified local constructors provide bounded receiver types, including inside macro token trees. Function signatures also supply parser-owned return types for local factory bindings through ?, unwrap/expect, and result-preserving error adapters. Instance factories returning Self resolve to their owning type, while single-argument generic wrappers propagate an inner receiver only when their signatures prove the same direct type parameter flows through. Recursive factory hints have a hard size budget. Unknown receiver types remain unresolved rather than being linked to an unrelated same-named method. forge plans every candidate byte, checks conflict baselines, validates the candidate in a copied workspace, runs Cargo build/tests when a manifest is present plus configured validation commands, and only then journal-commits the source projection and graph together.

collab exchanges semantic graph operations rather than text ranges. Each human or agent replica has a validated actor id and causal version vector; minimal idempotent deltas converge regardless of delivery order. Concurrent deletes win, concurrent updates have a deterministic tie-break, and deleting then recreating a node cannot resurrect edges from its old generation. RON .aetherc bundles are reviewable; .aethercb bundles use compact bincode. Init/sync reconciles source with the durable graph so graph-owned agent and extension metadata participates instead of being discarded. Bundle saves use a synced atomic replacement.

Membership is part of the causal operation history rather than a local address book. An approved collab fork registers the invited actor in both the source and forked bundles; if writing the fork fails, the source roster is rolled back. collab member add|remove ... --approve records convergent add/remove operations, concurrent removal wins, normal replica APIs reject new operations after the local actor is removed, and membership changes invalidate stale acknowledgements. A history has one genesis self-membership root: separately initialized actors cannot self-invite through a relayed delta. Removal records the highest counter observed for that actor; unseen later counters fail until their context observes a causal re-add, so a removed offline actor cannot keep extending a stale membership epoch. Version 1 and 2 bundles migrate conservatively by retaining the local actor, previously acknowledged peers, and non-bootstrap actors already present in the causal clock. Use fork to allocate a new actor replica; direct member add is for re-authorizing an already allocated unique actor, since it does not create that actor's bundle. Version 1-3 bundles load as explicit unsigned legacy history; version 4 stores operation attestations without changing existing CRDT dots.

Live host/join uses fresh random challenges, mutual HMAC-SHA256 group authentication, direction- and sequence-bound message integrity, bounded frames checked before allocation, and socket timeouts. Secrets are read from non-symlink regular files owned by the current user with private permissions. Optional identity mode adds transcript-bound Ed25519 proofs and a local actor-to-key trust store: both endpoints must configure it, each public key must match the peer actor's exact pinned fingerprint, and either attempted downgrade to group-secret-only mode is refused. The signed transcript also authenticates fresh X25519 keys, deriving a session-integrity key that another group-secret holder cannot calculate from captured traffic. Private identity files receive the same ownership, symlink, and permission checks. New trust, rotation, and removal are explicit fingerprint-approved operations, and identity files are actor-bound to their collaboration bundle.

Strict identity sessions also give every retained non-bootstrap CRDT operation a durable Ed25519 attestation. A delta importing a new operation, or a new retroactive attestation for an already-known dot, triggers verification of that actor's complete retained history before any bundle is replaced. The signature binds the dot, causal context, and exact action, so action tampering, actor forgery, unsigned relay, conflicting proofs, and replayed dot changes fail atomically. Local legacy operations can be upgraded with identity attest; identity verify audits a whole bundle against the local private identity and peer trust store.

Key rotation is a causal operation signed by the previous key with a second proof from the successor key. Verification walks this chain backward from the currently pinned fingerprint, so a retired key remains valid for its historical counters but cannot authorize later ones. Rotation operations are never removed by compaction. Deterministic bootstrap snapshot operations remain an explicitly pre-shared bundle baseline rather than pretending to have a human signature; strict deltas cannot introduce new bootstrap history. Group-secret-only live sessions and the unpinned collab merge workflow ignore incoming attestation metadata, preventing an unauthenticated path from poisoning later strict verification. Actor-key transparency beyond exact local pins is still a separate layer.

The transport deliberately binds loopback only: graph payloads are authenticated but not encrypted, so remote peers must connect through an encrypted tunnel such as SSH. After both sides verify that the other actor is active in the roster and durably persist a converged version, they persist monotonic peer acknowledgements. A delta that would remove either authenticated endpoint is rejected before persistence. A session claiming an unlisted actor is rejected even with a valid group-secret proof. Host and join may explicitly share a single-line status of at most 256 UTF-8 bytes. Both statuses and optional public keys are bound into the authenticated handshake. Status is reported to the peer and discarded after that synchronization; it is never written to graph operations, acknowledgements, discovery tickets, or collaboration bundles.

An optional --discovery-dir publishes an atomic, HMAC-authenticated, process-bound lease for the loopback host. The current-user directory and tickets must be private (new paths are mode 700 and 600 on Unix), symlinks are rejected, each scan is capped at 256 entries, dead process ids on Unix and actors outside the local bundle's active remote roster are ignored, and join-peer refuses ambiguous same-actor tickets. A normal host shutdown removes its unchanged ticket. Tickets are only endpoint hints: PID reuse or a stale ticket cannot authorize a session because the existing roster-bound mutual-authentication handshake still decides every join.

collab compact requires an acknowledgement from every active remote member, then prunes only causally superseded operations while retaining concurrent winners, membership removal barriers, and node-generation tombstones. Peers older than the recorded history floor fail safely and need a current bundle. Network/continuous discovery, continuous presence/subscriptions, encrypted remote transport, and operation-level signatures/key transparency remain future work. Group-secret-only migration mode remains a group credential: roster checks reject an unlisted claimed actor, but any secret holder can impersonate an active actor. Pinned identity mode prevents that endpoint impersonation after fingerprints have been verified. It does not retroactively prove the author of every historical CRDT operation: authenticated peers can relay the existing multi-actor operation set, so provenance of stored history remains trusted at the collaboration-group boundary until operations themselves are signed.

Every parsed module carries a bounded file-v1 whole-file projection in the semantic graph. collab review compares the remote and freshly reconciled local graphs, lists file and semantic changes, and reparses every remote file to prove its nodes and projection-derived edges agree with the claimed graph. Missing modules, path escapes, oversized files, inconsistent concurrent winners, and stale local baselines are conflicts. collab apply --approve reruns the plan, executes configured validation in an isolated candidate, then journal-commits added/modified/deleted files and the reconciled graph together. Native approval is SHA-256-bound to every candidate and expected baseline byte, so any project or bundle change forces another review.

Project configuration

Girder works without configuration. To materialize and inspect the validated defaults for a project:

girder config sample-project --init
girder config sample-project

girder.toml controls source roots and ignore globs, symlink policy, graph storage, structured impacted-test commands, the agent output module/file, and candidate-validation commands, timeouts, diagnostic limits, and copy budgets. Commands are argv arrays rather than interpolated shell strings. On Linux, available bubblewrap support mounts the host filesystem read-only while the candidate and approved build caches remain writable; other platforms still run inside the disposable candidate copy. Invalid keys, escaping paths, broken globs, unsupported graph extensions, malformed commands, and unsupported config versions fail before project analysis starts.

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 in
docs/core-gap-analysis.md) and refuses authored writes; run demos against
demo-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.

The pipeline demo

girder (no args) walks the entire pipeline over stdout:

1. Builds a semantic graph from source with tree-sitter (the graph is the source of truth; text is a projection).
2. Dispatches the agent swarm on a natural-language intent ("Add a multiply function to the math module"). The graph-aware Planner reads existing nodes before planning; the Coder generates each function, wires Calls edges, and emits a FeatureComplete summary; Tester, Documenter, Refactorer, SecurityAuditor, and Optimizer annotate in parallel.
3. Runs predictive impact analysis over the graph.
4. Time-travels a deliberate bug: records an execution trace via sys.settrace, branches a what-if alternative (variable overridden at a specific step via CPython's PyFrame_LocalsToFast), shows exactly where the two timelines diverge.

Workspace

The GUI

cargo run -p aether-app --features gui -- --gui sample-project

The native workspace opens a configured project, indexes its Rust, Python, and TypeScript source, provides file navigation and language-aware editing, folds edits into the graph, and reconciles those fresh projections with the durable graph. Agent summaries, graph-owned nodes, and inferred relationships survive reopen and editor refresh while source-derived structure follows the files on disk.

The semantic graph is an interactive navigation surface rather than a static diagram. Its retained force layout keeps positions stable while the graph changes; pan/zoom, fit-to-view, text search, node/edge type filters, one- and two-hop focus, viewport culling, and zoom-dependent labels keep large graphs legible. At overview scale, implementation nodes and their relationships collapse into weighted module-level connections; zooming in restores exact types, functions, and enabled relationships. The overview ranks and bounds module connections so the strongest architectural signals remain readable. Selecting a node exposes its source location and agent metadata, and double-clicking or choosing Open source moves the editor cursor to its span.

Editor saves use a recoverable journaled transaction for the source projection and semantic graph. Dirty buffers guard project/file switches, saves detect external source or graph changes, and startup rolls back an interrupted multi-file commit before indexing. GUI agent runs mutate a checkpointed graph and remain visibly pending while validation runs off the render thread. Commit stays disabled until the exact graph snapshot passes candidate build/tests; diagnostics remain inspectable and failed candidates can be revalidated or rolled back. Commit projects generated functions to the configured output file and persists the graph, while Roll back cancels validation and restores the pre-run graph without touching source files.

The right workspace has separate Agents, Extensions, Collaboration, and Author views. Extensions contains Generate, Marketplace, and Installed tabs. Extension generation returns a strict JSON recipe; installation stays disabled until the user reviews its exact SHA-256 digest, capabilities, contributions, projections, and full JSON. The marketplace searches bounded declarative catalogs, displays the catalog and listing/reference-recipe fingerprints plus reviews bound to both, and regenerates a listing intent against a bounded sample of the current semantic graph. Adapted recipes preserve the listing ID and show every added/removed capability scope. A catalog review never grants installation authority: the adapted recipe still requires a fresh approval bound to its own exact digest. Reviewer identities are catalog metadata rather than cryptographic signatures; verify an external catalog's printed SHA-256 fingerprint through the channel that distributed it.

The Collaboration view initializes, inspects, and synchronizes the full workspace graph, generates private secrets, verifies bounded local discovery tickets, and lets the user select an authenticated active-roster endpoint before joining on a background thread so rendering never blocks. It shows causal version/operation counts and the deterministic conflict policy, durable acknowledgements, and compacted history floor; it can conservatively compact acknowledged history. A live join updates the collaboration bundle only. Optional identity controls generate an actor-bound key, display a peer's exact fingerprint for out-of-band review, pin the unchanged public file, list trusted actors, and make strict identity mode visibly distinct from legacy group-secret mode. Explicit rotation/removal remains available in the CLI. Separate Review and Apply controls keep remote graph-to-source projection explicit, consistency-checked, digest-bound, validated, and atomic. CLI host is the persistent serving surface.

The Author view exposes both authoring paths from "External authoring" above without a terminal: type an intent, click Search to run the same concept search girder do/context use (only the top-scored hit starts checked — narrower than the CLI default on purpose, since gap 15 in docs/core-gap-analysis.md exists to shrink what a model can touch), and adjust the checkboxes to pin the exact nodes a plan may edit — the same --nodes a terminal invocation would otherwise require typing full paths for. Local model mode mirrors girder do: Run streams each provider attempt into a live log and shows pass/fail plus the report. External model mode mirrors girder context + plan run --authored: Copy context JSON puts exactly what the CLI command would print onto the clipboard, paste a model's plan response back in, and Run authored applies the same guarantees (forced rollback_plan, the mandatory tests.impacted check, zero-step refusal). A non-dry Run in either mode requires a second, explicit confirming click — a GUI button that silently writes to the tree has no command line to review first. Open demo-project/ to try the full loop for real:

cargo run -p aether-app --features gui -- --gui demo-project

Then, in the Author tab: type "add an exclamation mark to the farewell", click Search, leave crate::greeter::farewell checked, and either click Run (Local model) or use Copy Context JSON / paste a plan back in / Run authored (External model) — Dry run stays checked by default in both, so nothing writes to the tree until it's unchecked and confirmed. Like girder do against a terminal, both refuse sample-project/ outright; only demo-project/ (or another project you point it at) accepts a real, non-dry Run.

Installed records and their contribution nodes live in the semantic graph and survive source reconciliation. Enable/disable affects only contribution visibility. Removal conflict-checks every installed projection, restores replaced files, deletes files created by the extension, and commits the project plus graph as one recoverable transaction. Model output is never loaded as native code. Contributed validation commands resolve back to an installed, enabled recipe and require Bubblewrap; they run with networking disabled and a cleared environment in the disposable candidate workspace.

CI type-checks the gui feature on the stable Rust toolchain. Rendering needs a GPU — or a software Vulkan adapter (Mesa lavapipe) plus the usual X11 libs (e.g. libxkbcommon-x11). If no surface can be created the binary logs the wgpu error and falls back to the headless demo automatically, so it never hard-fails. It has been verified headlessly under Xvfb + lavapipe:

The default headless build, all tests, and girder's demo require none of this — no GPU, display, or extra system libraries.

Local-first AI

The default provider is a deterministic, offline MockProvider, so everything runs with no network and no API key. OpenAI and Anthropic are implemented live providers behind the live-providers feature. OpenAI uses the Responses API and is selected when OPENAI_API_KEY is set:

export OPENAI_API_KEY=sk-...
export OPENAI_MODEL=gpt-5.6   # optional; this is the default
cargo run -p aether-app --features live-providers -- forge sample-project "add a divide function"

Anthropic remains available as a second live provider:

export ANTHROPIC_API_KEY=sk-ant-...
cargo run -p aether-app --features live-providers -- forge sample-project "add a divide function"

The Router tries OpenAI first when configured, uses Anthropic as a secondary live provider for planning/codegen, and falls back to the mock otherwise. Gemini, xAI Grok, and local Ollama remain compile-clean extension-point structs; they deliberately do not enter routing until their HTTP bodies are implemented.

Status

Girder is a semantic-graph MCP server and CLI for coding agents. The native GUI, agent swarm, and time-travel debugger are supporting components; 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 measured recall 0.000 (docs/core-representative-mutations.md), which is why test selection is documented as advisory rather than authoritative.

Implemented features:

Optional DAP adapter smoke test:

python3 -m pip install debugpy
cargo test -p aether-dap --test debugpy -- --ignored --nocapture

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

Buy a license

Paid access to impacted_tests costs $39, one-time and perpetual, with no subscription. Buy a Girder license on Gumroad.

Set the purchased key as the complete value of the GIRDER_LICENSE_KEY environment variable. Alternatively, save it as the only contents of the key file that Girder reads for your platform:

• Linux and other non-macOS Unix: $XDG_CONFIG_HOME/girder/license.key, or $HOME/.config/girder/license.key when XDG_CONFIG_HOME is unset
• macOS: $HOME/Library/Application Support/girder/license.key
• Windows: %APPDATA%\girder\license.key

Verification is offline, and the key never expires.

License

Girder is source-available, not open source, under the Business Source License 1.1.

The free tier is genuinely free and permanent: it has no expiry and requires no account. For a single repository it includes get_source, find_definition, search_code, ask_codebase, review_changes, and orient. The impacted_tests tool requires a paid license.

Licenses are signed keys verified locally by the Girder binary. The binary never phones home, makes no network call for licensing, and works fully offline in both tiers.

On September 4, 2030, the license converts to the Apache License, Version 2.0. See LICENSE for the authoritative terms and CONTRIBUTING.md for contribution terms.