v2 Now with a HIR-driven workspace hypergraph & 45+ MCP tools

A workspace hypergraph,
for Rust agents.

An MCP server that combines hybrid BM25 + vector search with a HIR-driven workspace hypergraph built on rust-analyzer — exposing 45+ tools for symbol navigation, call-graph traversal, structural audits, and semantic neighborhood queries.

Get started View on GitHub Join Discord

45+// MCP tools

26// Claude skills

5// embed profiles

13// tool categories

// features

From keyword search to a queryable model of your workspace.

v2 layers a HIR-driven hypergraph on top of hybrid search — so your agent can audit structure, trace call relationships, and detect duplicate logic with the same query plane.

Hybrid search with RRF fusion

BM25 keyword matching fused with vector similarity via reciprocal-rank fusion. Find code by what it does, not just what it's called.

Persisted hypergraph

HIR-driven LMDB snapshot of cross-crate imports, exports, re-exports, call edges, attributes, signatures, statics, and unsafe blocks.

Pluggable embedding models

Local GPU (Qwen3 on Candle/CUDA), local CPU (BGE on ONNX), or API-backed (OpenRouter). New API models added via config — no recompile.

Safety & quality audits

unsafe_audit, mut_static_audit, recursion_check, channel-capacity, fn-body unwrap/panic/lock-across-await detection.

Call graph & structure

who_calls, calls_from, dead-pub detection, name collisions, forbidden cross-crate edges, Robert Martin instability metrics.

Codemap & semantic overlap

build_codemap produces task-conditioned Mermaid subgraphs; semantic_overlaps clusters duplicate logic across the workspace.

// tools

45+ MCP tools, grouped into 13 categories.

Plug the server into Claude Code (or any MCP client) and each tool becomes callable from your agent. Full parameter reference in TOOLS.md.

Query

search get_similar_code read_file_content

Symbol analysis

find_definition find_references rename_symbol get_dependencies get_call_graph analyze_complexity

Index lifecycle

index_codebase health_check clear_cache

Hypergraph build

build_hypergraph

Imports / exports

get_imports get_exports get_reexports get_declared_reexports

Reverse lookup

who_imports who_uses who_uses_summary

Call graph

who_calls calls_from call_graph callers_in_crate recursive_callers_count

Workspace structure

dead_pub_in_crate dead_pub_report crate_edges overlaps module_tree workspace_stats

Architecture rules

forbidden_dependency_check crate_dependency_metric

Signatures & attributes

function_signature functions_with_filter enum_variants item_attributes items_with_attribute

Safety & quality audits

unsafe_audit mut_static_audit recursion_check channel_capacity_audit fn_body_audit

Doc / API audits

missing_docs_audit derive_audit pub_use_pub_type_audit re_export_chain

Semantic & codemap

similar_to_item semantic_overlaps build_codemap

// claude code skills

26 ready-made audit recipes.

The skills/ directory composes these tools into self-contained Claude Code skills. Each is a SKILL.md with prerequisites, step-by-step prompts, and hand-offs to related skills. Install once and invoke as /<skill-name>.

Workspace · 7

rmc-workspace-overviewFirst-look audit of a Rust workspace rmc-architecture-rulesEnforce crate-edge rules rmc-dependency-metricRank by instability / abstractness rmc-imports-exportsCross-crate imports & exports rmc-type-overlapsName collisions, module shadows rmc-semantic-overlapsDuplicate logic by similarity rmc-snapshot-diffCompare two workspace snapshots

Crate · 3

rmc-crate-auditAudit one crate end-to-end rmc-api-surfaceAudit a crate's public API rmc-module-auditAudit a single module

Symbol · 7

rmc-find-symbolResolve a symbol's qualified name rmc-symbol-forensicsDeep dive: refs, callers, attrs rmc-method-apiAudit a type's methods rmc-trait-auditAudit a trait and its impls rmc-enum-variantsInspect an enum's variants rmc-rename-symbolPreview a rename — refactor probe rmc-signature-searchFind fns by signature shape

Quality · 7

rmc-unsafe-auditAudit unsafe blocks rmc-mut-static-auditAudit global mutable state rmc-attribute-auditAudit attributes & doc-comments rmc-complexityHotspots by blast radius rmc-test-vs-prodTest vs production split rmc-call-graphFn-level call graphs rmc-reexport-chainTrace pub use chains

Planning · 2

rmc-codemapTask-conditioned workspace subgraph rmc-refactor-planPlan a refactor with evidence

install skills

mkdir -p ~/.claude/skills
ln -s "$(pwd)/skills/"rmc-* ~/.claude/skills/

# then in Claude Code:
> /rmc-workspace-overview # kick off a tour of a new repo

// embedding models

Pick your quality / speed / privacy tradeoff.

Semantic search and the embedding-backed audits run on a configurable embedding profile. Switch by passing embedding_profile to the search tools; each profile gets its own independent index.

Profile	Model	Runtime	Dim
local-gpu-small default	Qwen3-Embedding-0.6B	local Candle / CUDA	1024
local-qwen3-4b	Qwen3-Embedding-4B	local Candle / CUDA	2560
local-qwen3-8b	Qwen3-Embedding-8B	local Candle / CUDA	4096
local-cpu-small	BGE-small-en-v1.5	local ONNX / CPU	384
openrouter-qwen3-8b	Qwen3-Embedding-8B	OpenRouter API	4096

embedding_profiles.toml

# add an API model without recompiling
[[profile]]
name     = "openrouter-text-embedding-3-small"
model_id = "openai/text-embedding-3-small"
dim      = 1536
max_len  = 8191

env

# OpenRouter keys come from the env, never config files
export OPENROUTER_API_KEY=sk-or-...

# local CPU / GPU profiles need no key

// installation

Build, wire it in, index.

Three steps to running. The Nix flake is the easiest path — it ships a dev shell with nightly Rust and the full CUDA toolchain.

1 · build

# clone and build the release binary
git clone https://github.com/molaco/rust-code-mcp.git
cd rust-code-mcp
cargo build --release

# or, with the Nix flake (recommended for CUDA)
nix develop github:molaco/rust-code-mcp
cargo build --release

# binary at target/release/rust-code-mcp

2 · .mcp.json

{
  "mcpServers": {
    "rust-code-mcp": {
      "command": "/abs/path/to/rust-code-mcp",
      "env": {
        "RUST_LOG": "info",
        "CUDA_HOME": "/usr/local/cuda",
        "LD_LIBRARY_PATH": "/usr/local/cuda/lib64"
      }
    }
  }
}

3 · index, build the hypergraph & query

# inside Claude Code:
> index_codebase at /abs/path/to/my-rust-project
> build_hypergraph # persisted in LMDB, fingerprinted, reused on next call

# then call any of 45+ tools — for example:
> unsafe_audit                  # list every unsafe block, with context
> dead_pub_report               # find dead pub items across crates
> crate_dependency_metric       # rank by instability / abstractness
> rename_symbol "parse_expr" -> "parse_expression"  # preview only
> build_codemap seeded by "main"  # Mermaid + outline subgraph

// architecture

One workspace, four indexes, one query plane.

A document index, a vector store, a HIR-driven hypergraph, and rust-analyzer's IDE engine — kept in sync by a Merkle-tree watcher.

01 · source

Rust workspace

Files on disk. Hashed via Merkle tree to detect change; re-sync every 5 min.

02 · text + vector

tantivy + lancedb

BM25 document index alongside an embedded vector store, fused with RRF.

03 · hypergraph

ra_ap_hir + heed (LMDB)

Persisted HIR snapshot: imports, exports, call edges, attrs, signatures, unsafe.

04 · surface

rmcp · MCP server

45+ tools exposed to any MCP client. 26 Claude Code skills compose them.

// performance

Indexing throughput, by profile.

Measured on the project's own repository (~2,280 chunks) with an RTX 3090. Embedding is ~95%+ of indexing time; vector search is exact brute-force KNN — fast at workspace scale regardless of dimension.

local-gpu-small

~55chunks/sec

Qwen3-0.6B on local GPU. Private, deterministic, no API calls.

openrouter · text-embedding-3-small

~220chunks/sec

Fastest profile. General-purpose model — not code-tuned.

openrouter-qwen3-8b

~45–50chunks/sec

API; per-request latency varies. Same model as local-qwen3-8b, no local GPU.

// stack

Standing on good shoulders.

Each piece is the best-in-class option for its job — and stays a thin wrapper over the upstream library.

tantivyfull-text BM25 fastembedembedding models candleGPU runtime (CUDA) reqwestOpenRouter client lancedbembedded vectors ra_ap_syntaxAST parsing ra_ap_idesemantic analysis ra_ap_hirhypergraph extraction heedLMDB graph store sledmetadata cache rmcpMCP protocol

MIT licensed · self-hosted · agent-friendly

Give your agent a real model of your codebase.

Clone the repo, plug it into Claude Code, and tell us how you use it — we want to hear about people's workflows.

github.com/molaco/rust-code-mcp Join the Discord

A workspace hypergraph, for Rust agents.

From keyword search to a queryable model of your workspace.

Hybrid search with RRF fusion

Persisted hypergraph

Pluggable embedding models

Safety & quality audits

Call graph & structure

Codemap & semantic overlap

45+ MCP tools, grouped into 13 categories.

Query

Symbol analysis

Index lifecycle

Hypergraph build

Imports / exports

Reverse lookup

Call graph

Workspace structure

Architecture rules

Signatures & attributes

Safety & quality audits

Doc / API audits

Semantic & codemap

26 ready-made audit recipes.

Pick your quality / speed / privacy tradeoff.

Build, wire it in, index.

One workspace, four indexes, one query plane.

Indexing throughput, by profile.

Standing on good shoulders.

Give your agent a real model of your codebase.

A workspace hypergraph,
for Rust agents.