semhood

Stop grepping. Find the exact code your AI agent needs by intent, not keywords. semhood is an AST-aware semantic code search engine that retrieves code by what it does, complete with call-graph context and optional LLM enrichment.

Runs fully offline with zero API keys — or plug in cloud embeddings — Voyage's code-specialized models, or OpenAI's strong general-purpose (natural-language) embeddings — for higher-quality retrieval. Optional LLM enrichment adds a logic summary and developer queries to each chunk that you commit once and share — and every result ships with its call graph (what it calls + what calls it).

Why semhood · Install · Quickstart · MCP Setup · Architecture · Config · Troubleshooting

   ┌──────────────┐    parse + embed     ┌──────────────┐    LLM enrich    ┌──────────────┐
   │  source tree │ ───────────────────► │ code  vector │ ──────────────► │  description │
   │              │                      │ + BM25 sparse│                  │  + queries   │
   └──────────────┘                      └──────────────┘                  └──────────────┘

                                Search via CLI, HTTP, or MCP.

semhood indexes your codebase the way a developer thinks about it: every function, method, and class becomes a chunk, with its call graph, docstring, and signature attached. You search by intent ("how do we retry transient payment failures?") and get back the few chunks that actually answer the question.

Two stages: structural (always, free, ~seconds) and enrichment (optional, LLM or code agent, drains a pending queue). The index is queryable after stage 1; stage 2 just makes natural-language matches sharper.

The package ships an MCP server so Claude Desktop, Cursor, Cline, Continue, Kiro, Zed, and any other MCP-aware client can call semhood as a tool — your AI agent gets search, find_symbol, and get_chunk_context next to its built-in read_file and grep.

What makes semhood different

Most code search embeds your raw source and hopes a natural-language query lands near it. semhood adds two things plain vector search and grep can't:

🧠 Enrichment that thinks like a developer

An LLM reads each function and class and writes two artifacts, each stored as its own search vector:

• A logic summary — what the code actually does, in plain language. Not its name, not its signature — its behavior.
• Predicted developer questions — the real questions an engineer would ask to find this code: "how do we retry a failed payment?", "where are customer emails validated?"

This is the part that makes retrieval click. When you search with a question, you're matching against questions the code was pre-labeled to answer — so the right chunk wins even when it shares zero keywords with your query. You generate the enrichment once, commit it to git, and your whole team — and every AI agent — retrieves better for free. → Portable enrichment

Why enrichment matters: some code has no words to search. Plain semantic search only works when the source contains language close to your question. Plenty of important code doesn't — terse names, raw math, business rules. Consider a function like:

def _calc(p, r, n):
    return p * (1 + r) ** n

Ask "how do we calculate compound interest?" and plain vector search comes up empty — there's no "interest", no "compound", nothing in the source that means anything close to the question. Enrichment reads the code and generates:

{
  "logic_summary": "Computes compound interest — final balance = principal × (1 + rate)^periods.",
  "developer_queries": [
    "how do we calculate compound interest?",
    "where is the future-value / compounding formula?",
    "how is a balance grown over multiple periods?"
  ]
}

Now your question matches a pre-written question that means the same thing, and _calc ranks first — despite sharing zero words with your query. The cryptic-but-critical functions are exactly the ones plain semantic search misses and enrichment rescues.

🕸️ Every result ships with its call graph

Each chunk knows what it calls and what calls it. So a result isn't just "here's the function" — it's the function plus its neighborhood. One get_chunk_context call returns the body, calls, and called_by together, so your AI agent gets the caller/callee context in the same response instead of opening files and tracing references by hand.

Features

• Retrieval-boosting enrichment — an LLM gives each chunk a plain-language logic summary + predicted developer questions, indexed as their own vectors so question-style queries hit the right code even with zero shared keywords. Written once, committed to git, shared with the team.
• Call graph included — each result knows what it calls and what calls it, so your agent gets the neighborhood, not just the function
• AST-aware chunking via Tree-sitter — every function, method, and class becomes a first-class searchable unit
• Hybrid retrieval — dense vectors (code, description, developer_queries) + BM25 sparse, all in one query
• Token-efficient for AI agents — replaces dozens of read_file + grep calls with one search
• MCP server out of the box — works with Claude Desktop, Cursor, Cline, Continue, Kiro, Zed
• 10 languages — Python, JavaScript, TypeScript, Java, Go, PHP, C#, Ruby, Rust, C++
• Free tier works — local embeddings + structural pipeline need zero API keys
• Incremental indexing — --changed re-indexes only what git diff touched

Install

# Local-only (free, offline embeddings via sentence-transformers)
pip install "semhood[local]"

# With everything: voyage + openai + anthropic + cohere + chroma
pip install "semhood[all]"

# Pick exactly what you need
pip install "semhood[anthropic,local]"

Prefer an isolated CLI install? pipx keeps semhood and its deps out of your global environment — recommended for a command-line tool:

pipx install "semhood[local]"

Just want to try it without installing? With uv, run it straight from PyPI:

uvx --from "semhood[local]" semhood index .

Requires Python 3.11+. First index downloads the embedding model (~420 MB for the default all-mpnet-base-v2) and caches it.

60-second quickstart

No config files, no API keys, no setup. semhood works offline with a local embedder by default. Just cd into any project and index it:

# 1. structural index — no LLM, free, no config needed
cd ~/code/your-project
semhood index .

# 2. search — pure retrieval, ~80 ms (a background daemon stays warm)
semhood search "how does the payment retry logic work?"

# 3. (optional) LLM enrichment — needs an Anthropic/OpenAI/OpenRouter/Ollama key
semhood enrich

# 4. (optional) full RAG with answer generation
semhood query "where is auth handled?"

The first command auto-starts a background daemon that loads the embedding model once and keeps it warm — so every later search (from any terminal or your editor) is instant. Each project gets its own index automatically under ~/.semhood/indexes/, keyed by repo root. One global config lives at ~/.semhood/config.yaml (created on first run); there is no per-project config file to manage.

CLI commands

All commands act on the current project (nearest git root of your cwd). Override with --root /path/to/repo. They're thin clients to the daemon — no model loading, no config.yaml flag.

MCP — wire it into your AI editor

The package ships an MCP server (semhood-mcp) that exposes eight tools — five for search, three for agent-driven enrichment:

Register it once, globally — it then works in every project you open, with no per-project setup. This is the entire config — drop it into your editor's user-level MCP file:

{
  "mcpServers": {
    "semhood": {
      "command": "semhood-mcp"
    }
  }
}

That's the whole config — no config.yaml path, no per-project entry. The server is a thin proxy: it forwards each call to the warm daemon, tagged with the project the editor currently has open (its nearest git root). Most editors launch the server with the workspace as the working directory, so this Just Works; if yours doesn't, pass the root explicitly:

{ "command": "semhood-mcp", "args": ["--root", "${workspaceFolder}"] }

After restarting the editor, the agent sees semhood.search, semhood.find_symbol, etc. The daemon loads the model once for the whole machine — so no matter how many editors and terminals you have open, there's one model in memory and every call is ~80 ms.

Teach the agent when to use it (optional)

The MCP server gives the agent the tools; a short skill teaches it to reach for semantic search before grep/read. Install it once, user-level, into every AI editor you use:

semhood install-skill            # auto-detects installed editors
semhood install-skill -t all     # or force every supported target

Supported: Claude Code (skill), Cursor & Windsurf (rules), Kiro (steering), Codex (AGENTS.md). It writes to each editor's own convention at the user level — so, like everything else here, it's set up once and applies to every project. Re-run any time to update in place.

Why this matters for token usage

Without semantic search, an AI agent asked "where do we validate emails?" runs grep -r email, gets 200 hits, and reads dozens of files. With semhood it calls search("validate email"), gets 3 ranked chunks back as JSON (~500 tokens), and reads only what it needs.

Worked example — "How does the OrderService validate orders?"

I indexed sample_project/, then asked exactly that. The agent calls two MCP tools and is done — no read_file, no grep.

{
  "query": "how does OrderService validate an order before processing",
  "embedding_dim": 768,
  "results": [
    {
      "score": 0.455,
      "qualified_name": "OrderService::_validate",
      "file": "orders.py", "line_start": 115, "line_end": 124,
      "type": "method", "enrichment_state": "pending"
    },
    {
      "score": 0.397,
      "qualified_name": "OrderService",
      "file": "orders.py", "line_start": 49, "line_end": 124,
      "type": "class",
      "summary": "High-level orchestrator for placing and processing orders. Coordinates payment authorization (PaymentProcessor) and customer notification (NotificationService)."
    },
    {
      "score": 0.363,
      "qualified_name": "OrderService::place_order",
      "file": "orders.py", "line_start": 67, "line_end": 93,
      "summary": "Validate, charge, and confirm an order in one shot."
    }
  ]
}

{
  "found": true,
  "chunk": {
    "qualified_name": "OrderService::_validate",
    "file": "orders.py", "line_start": 115, "line_end": 124,
    "source": "def _validate(self, order: Order) -> None:\n    if order.is_empty():\n        raise ValueError(\"Order has no line items\")\n    if order.total_cents() < self.MIN_ORDER_CENTS:\n        raise ValueError(...)\n    if \"@\" not in order.customer_email:\n        raise ValueError(\"customer_email must be a valid email address\")",
    "calls": ["order.is_empty", "ValueError", "order.total_cents"],
    "called_by": [
      {"function": "place_order", "class_name": "OrderService", "file": "orders.py", "line": 67}
    ]
  }
}

Compared to the grep-then-read alternative — open orders.py (3.5 KB), scan for validate, then trace into place_order to confirm the call — the MCP path uses roughly 40% of the tokens and skips file I/O entirely.

Two ways to enrich

Stage 2 (description / developer_queries vectors) needs an LLM. You pick how:

Option A — direct API

semhood calls Claude / GPT / OpenRouter / Ollama itself.

semhood enrich

Best for CI, batch jobs, or users without an AI editor. Set llm.provider + llm.<provider>.api_key in config.yaml.

Option B — your AI agent does it

Skip the API key. The agent in your editor (Claude in Cursor, Kiro, Cline, Continue, etc.) loops over pending chunks via the MCP tools and writes summaries itself. No semhood-side LLM cost. Just say:

"enrich the codebase index"

The agent uses list_pending_enrichments → writes summary → save_enrichment → repeats until enrichment_progress reports zero pending. Uses the model + subscription you already pay for.

Portable enrichment — pay once, share, survive model changes

Enrichment (the LLM-written summaries + developer queries) is the only part of semhood that costs tokens. semhood persists that text to .semhood/enrichment.jsonl in your repo, keyed by each chunk's content hash — separate from the vector index (which lives in ~/.semhood/indexes/ and is disposable).

Commit .semhood/enrichment.jsonl to git. Doing so gives you:

• Share across a team — a teammate clones, runs semhood index, and the enrichment is restored from the file and re-embedded locally. They pay zero LLM tokens for code that's already enriched.
• Survive embedding-model changes — switch from a local model to Voyage/OpenAI, run semhood index --reset, and your enrichment is re-embedded into the new vector space with no re-enrichment. (Without this, changing models meant re-paying for everything.)
• Survive branch switches / reverts — old records are kept as a content-addressed cache, so checking out an older revision reuses its enrichment instantly.

It's content-addressed, so it's always safe: if a chunk's code changed, its hash misses and that chunk is simply re-enriched — stale summaries can never attach to changed code. Run semhood compact to prune records no longer referenced by any code. The file holds only text and contains no secrets.

.semhood/enrichment.jsonl   ← commit this (portable enrichment, ~text)
~/.semhood/indexes/<id>/    ← never committed (vectors, machine-local)

OpenRouter — one key, every model

If you don't want to manage Anthropic + OpenAI + Google accounts separately, point semhood at OpenRouter:

llm:
  provider: "openrouter"
  openrouter:
    api_key: "${OPENROUTER_API_KEY}"
    model: "anthropic/claude-sonnet-4.6"   # or openai/gpt-5.5, anthropic/claude-opus-4.7, etc.

Use any supported model slug. Same semhood enrich and semhood query commands as before.

Architecture

Three named dense vectors per chunk: code (always), description (post-enrichment), developer_queries (post-enrichment), plus a bm25 sparse vector. Pre-enrichment, search still works on code + BM25 alone.

Stage 1 (semhood index) parses files with Tree-sitter, builds the call graph, computes a source hash, embeds the raw source into the code vector, and upserts to Qdrant. No LLM calls. Every chunk lands in state pending.

Stage 2 (semhood enrich) drains the pending queue through an LLM, generating a logic_summary + 5–10 developer_queries per chunk, embedding them into the description and developer_queries vectors and mirroring the text to the portable .semhood/enrichment.jsonl store. State flips to fresh. Trivial chunks get a free template; tests get pattern-skipped.

Configuration

semhood uses one global config at ~/.semhood/config.yaml (created with working defaults on first run) — the daemon loads it for every project. You can still point at an explicit file with $SEMHOOD_CONFIG. Environment-variable refs (${VAR}) resolve from ~/.semhood/.env, a local .env.semhood/.env, or the real environment. Indexes are stored centrally under ~/.semhood/indexes/<id>/, one per project root — you don't set a path.

embeddings:
  provider: "local"            # local | voyage | openai | cohere
  local:
    model: "sentence-transformers/all-mpnet-base-v2"
  # voyage:                    # code-tuned, no local load — needs an API key
  #   api_key: "${VOYAGE_API_KEY}"
  #   model: "voyage-code-2"

llm:
  provider: "anthropic"
  anthropic:
    api_key: "${ANTHROPIC_API_KEY}"
    model: "claude-sonnet-4-6"

vector_db:
  provider: "qdrant"        # path is managed per-project by the daemon

enrichment:
  enabled: true
  version: 1
  llm_concurrency: 4
  skip_patterns: ["**/test_*.py", "**/*_test.go", "**/tests/**"]
  skip_trivial:
    enabled: true
    max_lines: 3

Supported languages

Troubleshooting

Status

Tune enrichment.llm_concurrency, or use the agent-driven enrichment flow if you hit rate limits.

Contributing

Issues and PRs welcome.

# Run tests
pytest

# Build a release
python -m build

See open issues for things to pick up.