seam-index 1.0.2

Local-first semantic codebase index for AI agents

Seam

Your codebase, indexed once. Queried by any agent, instantly.

Seam is a local-first CLI and MCP server that gives coding agents semantic access to a repository without re-reading the whole codebase on every task.

Install

Linux/macOS curl installer

Install from GitHub using the auto installer:

curl -fsSL https://seam.adityamer.dev/install.sh | bash

Install the CLI plus the AI agent skill:

curl -fsSL https://seam.adityamer.dev/install.sh | bash -s -- --with-skills

Installer options:

# Prefer PyPI once the package is published
curl -fsSL https://seam.adityamer.dev/install.sh | bash -s -- --source pypi

# Force a specific installer backend
curl -fsSL https://seam.adityamer.dev/install.sh | bash -s -- --method uv
curl -fsSL https://seam.adityamer.dev/install.sh | bash -s -- --method pipx
curl -fsSL https://seam.adityamer.dev/install.sh | bash -s -- --method pip

Python installers

uv tool install git+https://github.com/Aditya190803/seam.git
pipx install git+https://github.com/Aditya190803/seam.git
python3 -m pip install --user git+https://github.com/Aditya190803/seam.git

After PyPI release:

uv tool install seam-index
pipx install seam-index
python3 -m pip install --user seam-index

AI agent skill installer

Install the portable Seam skill through skills.sh for supported AI coding agents:

npx @aditya190803/seam-skill install

Project-local skill install:

npx @aditya190803/seam-skill install --project

By default this delegates to npx skills add and targets all skills.sh-supported agents. Use --agent <name> to target specific agents, e.g. --agent claude-code --agent pi.

Local development

uv sync
uv run seam --help

The package exposes a seam console command.

Quick start

# Index the current repo
uv run seam init .

# Search with human-readable output
uv run seam search "where is JWT validation handled"

# Search with JSON output for agents
uv run seam search --json --top 5 "database connection pooling"

# Generate ready-to-paste context
uv run seam context --format xml "database connection pooling"

# Show index metadata
uv run seam status

Indexes and config are stored under ~/.seam by default. Set SEAM_HOME=/path/to/home to override this location.

Agent skill

Seam ships a portable Agent Skill at skills/seam-code-search/SKILL.md. It works with tools that support the Agent Skills format:

• Codex discovers repo skills in .agents/skills/ and user skills in ~/.agents/skills/.
• Claude Code discovers repo skills in .claude/skills/ and user skills in ~/.claude/skills/.

Install it with npx or the Seam CLI:

npx @aditya190803/seam-skill install
seam install-skill

See docs/agent-skills.md for details. Once installed, ask an agent to “use Seam to find the relevant code” or invoke the skill directly if your agent supports explicit skill invocation.

Commands

seam init [path]              Index a codebase; use --warm to warm local index files
seam reindex [path]           Re-index a repository, file, or directory subtree
seam search [opts] <query>    Search indexed code semantically or with hybrid ranking
seam grep [opts] <pattern>    Exact literal/regex search over indexed chunks
seam context [opts] <query>   Generate markdown/xml/json context blocks
seam status [--size]          Show index freshness, metadata, and optional disk size
seam gc                       Remove index entries for files no longer on disk
seam check                    Script-friendly health check
seam watch [path]             Watch files and refresh the index
seam serve                    Start MCP server on stdio transport
seam serve --http <port>      Start MCP server over HTTP when supported by FastMCP
seam export <archive.tar.gz>  Export a local index archive
seam import <archive.tar.gz>  Import a local index archive
seam doctor                   Check dependencies and service reachability
seam install-skill            Install the Seam Code Search Agent Skill
seam config show|list         Print current configuration
seam config set <key> <value> Set a configuration value

Useful search options include --path (repeatable), --all-repos, --language, --name, --exclude, --mode hybrid|semantic|keyword, --exact, --changed, --count, --min-score, and --alpha.

JSON search output is an object with duration_ms and results; see docs/seam.schema.json for the machine-readable schema.

Current implementation

This repository implements the Seam product shape with production backends where they can be used locally:

• Typer CLI for the PRD command surface
• Offline deterministic local embeddings by default
• OpenAI-compatible and Ollama embedding providers
• SQLite local vector store
• LanceDB local vector backend
• Qdrant remote vector backend using metadata-only payloads; source snippets stay in the local Seam metadata store and local repo
• Incremental file hashing, scoped reindexing, deletion detection, .seamignore glob exclusions, and persisted Merkle tree metadata
• Optional tree-sitter chunking through tree-sitter-language-pack, with Python AST and structural fallbacks
• Hybrid vector + keyword ranking with filename/exclusion filters, exact grep-style search, all-repo search, and changed-file filtering
• Watchdog-based seam watch with localhost:7731/health
• FastMCP tools: search_code, list_files, get_chunk, index_status
• seam export / seam import for portable local index archives

The default is local SQLite + deterministic embeddings for zero-config, offline operation.

Configuration

seam config show
seam config set embedding_provider local
seam config set backend sqlite   # or lancedb / qdrant
seam config set hybrid_search true

OpenAI-compatible embeddings can be selected with:

export OPENAI_API_KEY=...
seam config set embedding_provider openai
seam config set embedding_model text-embedding-3-small
seam reindex .

Qdrant can be selected with:

seam config set backend qdrant
seam config set qdrant_url http://localhost:6333
seam config set qdrant_api_key_env QDRANT_API_KEY
seam reindex .

Ollama embeddings can be selected with:

seam config set embedding_provider ollama
seam config set embedding_model nomic-embed-text
seam config set embedding_base_url http://localhost:11434
seam reindex .

MCP

seam serve starts FastMCP on stdio, which is the most portable transport for local agents. seam serve --http 7732 starts HTTP/streamable HTTP when supported by the installed FastMCP version. Unix socket transport from the PRD is not exposed directly by this FastMCP release; use stdio for local agent integrations.

More docs

• Install guide
• Agent skill guide
• Benchmarks
• Release checklist
• Security policy
• Contributing

Development

uv run ruff check .
uv run pytest -q
uv run python -m compileall seam main.py scripts/benchmark.py