ai-architect-mcp-codebase parses a repository with tree-sitter, stores the
symbols and their relationships (definitions, calls, imports, implementations,
uses) in an embedded LadybugDB graph database on your machine, and exposes
that graph to an agent through MCP tools. An agent can search the graph, look
up a symbol with its callers and callees, list the reverse dependencies of a
symbol, follow execution flows from entry points, map a git diff onto the
symbols it changes, and compare two graphs of the same code. Answers about
symbols carry source locations and the limits of the analysis behind them.
Codebase intelligence MCP server — index any repo into a property graph; 26 tools for agents.
2026-11-09
스타
—
도구
—
가동 시간
—
The server does not edit the files it indexes. It writes its graph and
sidecar files under the output_dir you pass, plus a snapshot in the
repository only when you ask for the team-shared artifact. Indexing and queries
run locally without network calls, no language model reads the code, and
nothing is uploaded. The graph is a directory you own.
The graph comes from static parsing of source text, so some things stay out
of it. A call made through reflection, dynamic dispatch, generated code, or a
macro the parser does not expand has no edge unless an optional
language-server pass (lsp: true) or runtime traces (ingest_traces) supply
one. Method calls through a receiver are resolved statically only for the
receiver shapes listed below; the rest need
the language-server pass. Ruby gets a shallow extraction with no imports or
visibility. Process flows are reachability in the graph from declared entry
points, capped at depth 20, with no observation of execution. An empty answer
can mean "no such relationship" or "not indexed", and the coverage report
exists to tell those apart.
Install and register
Claude Code plugin
claude plugin marketplace add cdeust/ai-architect-mcp-codebase
claude plugin install ai-architect-mcp-codebase@ai-architect-mcp-codebase-marketplace
The plugin downloads the release binary and verifies its Sigstore provenance
before running it; this needs GitHub CLI 2.68 or newer. The plugin starts the
full profile unless AP_PROFILE=core is set.
Other MCP hosts
cargo install ai-architect-mcp-codebase
ai-architect-mcp-codebase install # writes the MCP entry for each detected host
install detects Claude Code, Codex CLI, Gemini CLI, Cursor, VS Code and Zed,
adds only its own entry to each config file, and never overwrites a file it
cannot parse. --dry-run previews the changes; uninstall removes them. To
configure a host by hand, register the binary as a stdio server:
docs/install.md has the per-host snippets (Codex, Gemini,
Cursor, Windsurf, VS Code, the OpenAI Agents SDK), the optional Grep/Glob
hook, the plugin's release verification, and the developer escape hatch for
running a local build inside the plugin.
Build from source
git clone https://github.com/cdeust/ai-architect-mcp-codebase.git
cd ai-architect-mcp-codebase
cargo build --release # about five minutes the first time (LadybugDB's C++ core)
Recommended for agents. Analyze once, then search, inspect and measure impact.
full
all 27
Adds the manual graph passes, history, runtime traces, the verification tools, the artifact builder and the finding-record tools.
ai-architect-mcp-codebase --profile core # the 8 core tools
AP_PROFILE=core ai-architect-mcp-codebase # same, through the environment
ai-architect-mcp-codebase # default: full (all 27)
The default stays full until the next major version, because shrinking the
default tool surface is a breaking change. New agent setups should use the
core profile (8 read-only tools): analyze_codebase already runs index,
resolve and cluster in one call, and leaving out the 19 hidden tools keeps the
tool prompt small.
A first session
health_check()
→ server name and version, protocol version, registered tool count
analyze_codebase(path: "/path/to/repo", output_dir: "/tmp/repo-graph")
→ index, resolve, cluster and build the search index in one call
→ node and edge counts, communities, processes, and the coverage report
search_codebase(graph_path: "/tmp/repo-graph/graph", query: "tool request dispatch")
→ ranked symbols (BM25 and sparse TF-IDF fused with RRF), paged with next_offset
get_context(graph_path, qualified_name: "src/main.rs::handle_tool_call")
→ the symbol with its callers, callees, imports, implementations,
community and processes; suggestions when the name does not match exactly
get_impact(graph_path, qualified_name: "src/main.rs::handle_tool_call")
→ callers, importers, users and implementors, the communities and processes
affected, and an epistemic label (exact or lower-bound) with its reasons
detect_changes(graph_path, codebase_path: "/path/to/repo", base_ref: "main", head_ref: "HEAD")
→ symbols the diff adds, modifies or deletes, with a heuristic risk score each
query_graph(graph_path, graph: "missed")
→ the files the index could not fully cover, to check with grep
Re-run analyze_codebase after edits; indexing is incremental by default.
27 MCP Tools
Each tool takes JSON arguments checked against its JSON Schema and returns
structured JSON, with a named reason code on error. No tool calls a language
model. Agent installs rarely need all 27; the core profile registers just
the 8 code-intelligence tools.
Returns server name and version, MCP protocol version and the registered tool count, which shows the active profile.
analyze_codebase
yes
Runs index, resolve and cluster in one call and builds the search index; lsp: true adds the language-server pass. Returns statistics and the coverage report.
index_codebase
Walks and parses the repository into <output_dir>/graph/. Incremental by default; can export or import the team-shared graph artifact.
resolve_graph
Adds cross-file Imports, Calls, Implements, Extends and Uses edges with confidence scores.
lsp_resolve
Resolves remaining calls through rust-analyzer, pyright or typescript-language-server.
cluster_graph
Detects communities (Louvain with C2 repair) and traces processes from entry points.
ingest_traces
Folds runtime caller-to-callee observations (OTel spans, profiler or coverage traces) into the graph, and adds OBSERVED_CALLS edges where static resolution found none.
index_status
Node and edge counts plus the coverage report.
search_codebase
yes
Hybrid ranked search over symbols, and BM25 search over documentation files.
get_symbol
yes
Exact lookup by qualified name with every incoming and outgoing edge; sibling_graphs looks for a definition in other repositories' graphs.
get_context
yes
One symbol with its relationships grouped by kind, its community and its processes.
query_graph
yes
Read-only Cypher, paged; graph: "missed" lists what the index did not cover.
get_impact
yes
Reverse dependencies of a symbol or file, with affected communities and processes and an epistemic label.
get_processes
Lists execution flows with entry point, entry kind (main, test, proof, handler, lib_entry), depth and size.
detect_changes
yes
Maps a unified diff, or base_ref..head_ref, onto the symbols, communities and processes it touches.
index_history
Adds git history as Commit and Version nodes linked to the files and symbols each commit changed.
check_security_gates
Checks a list of changed symbols for auth-community touch, public-API change, unresolved imports and test-coverage gaps.
verify_semantic_diff
Compares a before graph and an after graph: added and removed nodes and edges, dangling references, new unresolved imports, new cycles (Tarjan SCC), and a heuristic regression score.
validate_prd_against_graph
Checks a product requirements document against the graph: symbols it names that do not exist, changes that span several communities, and "does not affect X" claims contradicted by process membership.
check_doc_claims
Checks what a document claims about the code, one claim at a time, each anchored to the exact text at a file:line: symbols, files and modules that exist, items declared pub, test and proof counts, enum variant counts. Every row is supported, contradicted, not found, not verifiable (with the reason) or rejected for a bad anchor, with a replayable query as evidence. An absence, or a count the graph holds only as a lower bound, is never called contradicted.
prepare_prd_input
Writes a JSON bundle of graph facts (matched symbols, affected communities and processes, graph statistics) for a finding or a free-text feature description.
extract_finding, refine_finding
Normalize an incoming finding to a canonical JSON record and store an agent's refinement of it.
Record a question-and-answer clarification of a finding, finalized with a SHA-256 digest of the transcript.
validate_prd_against_graph, prepare_prd_input and the finding-record tools
produce or check JSON artifacts on disk under the output_dir a call names. A specification tool such as
ai-architect-mcp-spec, or any other consumer, can read
them; the server itself has no dependency on the consumer.
The verification verdicts have narrow meanings. check_security_gates returns
gates_passed: true when no check raised a critical flag, and
report.assessment_complete separately; the latter is false for an empty
symbol list, a skipped check, or changed symbols that could not be resolved.
The unsafe-symbol check is skipped until the Rust parser records unsafe.
Neither field certifies security. A clean verdict from
verify_semantic_diff needs a regression score below the threshold and no new
unresolved import, and says nothing about behavioural equivalence or
compilation; tests, the compiler or proofs establish those. A finalized
clarification digest binds the recorded transcript bytes and does not
establish that the finding is true.
How answers state their limits
The read tools report what they could not see next to what they found. The
0.12.0 entry of the CHANGELOG has the detail for the coverage,
impact, language-server, receiver-resolution and freshness items below.
Coverage report
analyze_codebase, index_status and query_graph(graph: "missed") report
files that were parse_incomplete, skipped or quarantined, plus three
Rust buckets added in 0.12.0. outside_build_targets lists indexed .rs
files outside every compiled Cargo target according to
cargo metadata --no-deps (a Kani proof file, an excluded fuzz/ directory).
feature_gated lists modules reached only through a mod declaration whose
#[cfg(feature = ...)] is false under default features. unlinked_file lists
files rust-analyzer itself reports as belonging to no crate, recorded when an
LSP pass opens them. cargo_attribution.status says whether the two
Cargo-derived buckets were determined: known, unknown (the cargo metadata call failed or cargo is missing, so an empty bucket means "not
determined"), not_applicable, or not_recorded for an older record.
coverage.pruned_dirs names every directory the walk skipped by policy, with
its reason. An empty coverage report is still not proof of completeness.
Epistemic boundary on impact
get_impact labels its answer exact only when the coverage record has zero
gaps in every bucket, cargo_attribution.status is not unknown, and no
unresolved call site names the target. Otherwise the answer is lower-bound, with reasons.
unresolved_callsites_naming_target counts unresolved call sites whose callee
names the target, and unresolved_callsites_outside_targets counts those in
files outside the compiled Cargo targets, so an empty caller list can be told
apart from a symbol with no callers.
Language-server pass
With lsp: true, analyze_codebase reports lsp_status.state as disabled,
completed, completed_unresolved (sites needed resolving and none were
resolved) or failed. A Rust crate nested under a Cargo workspace that does
not list it as a member fails the pass with lsp_workspace_load_failed, naming
the Cargo-level fix, before any resolution request is sent. Analysis continues
on the available graph when the pass fails. server_health carries the language
server's own last-reported health.
Static receiver-call resolution
Since 0.12.0 the static resolver binds self.m() and Self::m() in Rust,
self.m() in Python and this.m() in TypeScript to the enclosing type's
method (resolution_method: "receiver-type"). In Rust it also binds x.m()
when x is bound once in the enclosing function by a typed parameter, a typed
let, or let x = T::assoc(..) ("receiver-local-binding"). It also
types x from the declared return type of a free function of the same file
that initialised it (let x = make();, and an Option or Result unwrapped by
let Some(x) = .. else, .expect(..), .unwrap() or ?), at a lower
confidence ("receiver-return-type", 0.85). A receiver it cannot type is left
unresolved; it never falls back to a lookup by bare name.
Query truncation and paging
query_graph bounds a query with no LIMIT of its own to 500 rows per page
(limit_injected: true). truncated is true whenever more rows exist after
the page, and next_offset reads the next one. When the row bound cut the
result, total_count is a lower bound. Paging is stable only when the query
declares ORDER BY, which order_stable reports. The tool's JSON Schema
description has the full contract.
Freshness receipt
search_codebase, get_symbol and get_impact return graph_freshness
(state: fresh | stale | unknown, dirty_files, checked_files,
commits_behind, commits_ahead) on every response, errors included, so an
answer from a graph older than the working tree is marked as such. A graph
indexed by 0.11.x reads unknown until it is re-indexed.
Languages
Rust, Python, TypeScript (and JavaScript through the TSX grammar), Java,
Kotlin, Swift, Objective-C, C, C++ and Go get deep extraction through a
per-language spec: definitions, calls and imports, plus visibility and type
relationships where the spec records them. Ruby is a shallow tier
(ADR-0056): functions,
methods, classes and modules, and calls, with no visibility and no import
edges (require appears as a call). Other files, documentation and
configuration included, are indexed as File nodes. The text of documentation
files (md, markdown, mdx, txt, rst, adoc, up to 256 KiB) is
searchable through BM25. The language-server pass supports Rust, Python and TypeScript.
Security model
The server enforces read-only access to the graph in query_graph with two
layers. The engine's compiled-plan check refuses database writes, and a
lexical gate refuses the filesystem statements the engine classifies as
read-only (COPY … TO, EXPORT DATABASE, ATTACH, and others), plus every
CALL outside a two-procedure allowlist. Git refs are validated, the
language-server command is limited to four known binaries, the walk has size
and depth limits, and parsing has a per-file deadline. Release binaries carry
Sigstore provenance that the Claude Code plugin verifies before running them.
The dominant energy cost in an LLM-assisted coding workflow is the model
inference spent re-reading files to answer a question a structural query could
answer once; this binary's CPU is small next to it. Every token an agent does
not process is compute that is never scheduled. The
Green Software Foundation calls this
energy proportionality, and here it is applied where the constant is largest.
The head-to-head evaluation measures that demand under a pre-registered
protocol, n=20, reproducible offline with no API key:
Demand term
Graph tools
Grep/Glob/Read baseline
Reduction
Payload token proxy
43.14 ± 17.26
550.36 ± 330.28
14.26x (mean of per-question ratios)
Modeled tool calls
1.00 ± 0.00
5.20 ± 1.64
5.20x
Costs are modeled. Both legs use a payload-size / 4 token proxy, and indexing,
client prompts, real MCP response envelopes and model reasoning are excluded.
The corpus informed fixes #87 and #92, so this is a regression benchmark.
Dividing aggregate payload volumes gives 12.76x, a different statistic. No
energy or CO2 figure is derived from these numbers, because this repository
measures no joules and a token count does not convert to watt-hours.
docs/evaluation.md
has the full table, the protocol and the falsified first run.
On the server side, bulk graph writes through UNWIND with a typed struct cost
0.127 ms per edge against 9.658 ms for the naive raw-string path, 76 times
less work for the same output (measured 2026-07-28 on lbug 0.18 with
tests/lbug_bulk_investigation.rs). Tree-sitter extracts structure without a
model, and indexing is incremental by default. The server-side measurements
and their dates are in docs/evaluation.md.
Every algorithm traces to a source. Louvain: Blondel et al. 2008. Leiden
C2 repair: Traag et al. 2019. RRF: Cormack, Clarke and Büttcher 2009.
Strongly connected components: Tarjan 1972. BM25 (through Tantivy):
Robertson et al. 1994.
Named constants record their source or measured rationale. The RRF
constant K = 60.0 in src/search/rrf.rs cites Cormack 2009;
BULK_BATCH_SIZE = 500 cites LadybugDB practitioner guidance and the April
2026 scalability audit. CONTRIBUTING.md requires a
// source: annotation on every numeric constant with three or more
significant digits, and an efficiency claim with no measurement behind it
does not ship.
No invented numbers. Where a value was chosen by judgment, the comment
says so and gives its operational justification.
Error responses name the rule that was violated, for example
unsafe finding_id (spec §5.1.4, §9.3 Q4): must not contain '..'.
When a capability is unavailable, the tool says so in plain language. For
example, lsp_resolve on a binary that is not a language server returns
lsp_probe_failed: found on PATH but didn't respond as an LSP server (stdout closed immediately; likely a stub, proxy, or non-LSP binary).
Development
cargo build --release
cargo test # full suite (3200+ tests)
cargo test --test graph_accuracy # structural accuracy gate
cargo clippy --all-targets -- -D warnings # zero warnings, enforced in CI
python3 scripts/check_doc_claims.py # README numbers against their sources
CI also measures line coverage with cargo llvm-cov (80% floor) and checks the
coverage and test-count badges above against that run. Contributions need a
DCO sign-off (git commit -s); CONTRIBUTING.md has the
layer rules and coding standards, and
docs/architecture.md the module map, dependencies and
design records. The evaluation programs are described in
docs/evaluation.md.
Repository layout
ai-architect-mcp-codebase/
├── src/
│ ├── main.rs ← MCP server entry point and dispatch
│ ├── tool_schemas.rs ← JSON Schemas, one file per tool group
│ ├── tool_profile.rs ← core and full profiles
│ ├── graph_store/ ← LadybugDB port
│ ├── parser/ ← tree-sitter extraction, one spec per language
│ ├── indexer/ ← walk, parse, persist, coverage
│ ├── resolver/ ← cross-file resolution
│ ├── lsp_client/ · lsp_resolver/
│ ├── clustering/ ← communities, processes, impact
│ ├── search/ ← BM25, sparse TF-IDF, RRF
│ ├── history/ ← git history layer
│ ├── prd_input/ · prd_validator/ · security_gates/
│ └── semantic_diff.rs · git_diff.rs
├── tests/ ← integration tests and graph_accuracy.rs
├── benchmarks/ ← eval_headtohead, token_surface, incremental_speed
├── stages/ ← design specs and decision records
├── docs/ ← install, configuration, security, architecture, evaluation
├── scripts/ ← doc-claim and pin gates
├── bin/ ← plugin launcher and release verification
├── skills/ ← host-native workflows
├── plugins/ ← Codex plugin package
├── mcp-contract.json
└── Cargo.toml
Status
Version 0.14.0. The capabilities above are covered by the test suite and the
accuracy gate in CI and have been run end to end on the maintainer's machine;
they have not yet been validated in a production deployment. The unsafe-symbol
security check waits on unsafe extraction in the Rust parser. Rename and
refactor tools are out of scope because the server does not edit code. Direction is in
docs/ROADMAP.md; what shipped is in
CHANGELOG.md.
Related projects
Each of these runs on its own. Cortex is
a persistent-memory MCP server, and
ai-architect-mcp-spec is an
MCP server for PRD generation and specification verification. An agent that
has both this server and ai-architect-mcp-spec registered can ground a
specification in the graph through search_codebase, get_context and
get_impact, or through the JSON file prepare_prd_input writes, and check
the result with validate_prd_against_graph.
This software is the independent work of Clément Deust. It was developed
outside any employment relationship and is not affiliated with, endorsed by,
or owned by any past or present employer.
The graph and retrieval algorithms used here (Louvain community detection with
C2 repair, BM25, RRF rank fusion, tree-sitter parsing, Tarjan strongly
connected components) come from published research; citations are inline in
// source: annotations and in docs/. The MIT license covers this
implementation and asserts no ownership over the underlying algorithms, which
remain attributable to their original authors.