code-index-pg

Unified code index for Hermes: exact + semantic + call graph search over the TGT One monorepo, persisted in postgres. Replaces the code-index fork, cocoindex and CodeGraphContext with one lightweight MCP server.

Stack

postgres-local (port 5432, DB codeindex) + pgvector 0.8.3
Ollama bge-m3 (1024 dims) for embeddings — no torch in this venv
tree-sitter-language-pack (21 languages) for symbol extraction
fastmcp — single MCP process, no extra daemons

Install

git clone https://github.com/tgt-technology/code-index-pg.git
cd code-index-pg
uv sync
uv tool install --editable .

DB: la infraestructura (postgres + pgvector, ollama + bge-m3, base codeindex) se levanta con deploy/ — ver Deploy. El esquema de tablas lo aplica el server al arrancar.

Deploy (entorno desde cero)

El índice vive en postgres, no en archivos: reproducir el entorno es levantar la infra, no copiar datos. Todo está en deploy/.

Windows (Docker Desktop + uv):

powershell -NoProfile -File deploy\setup.ps1        # -Gpu si hay GPU NVIDIA

Linux / WSL:

deploy/setup.sh          # --gpu si hay GPU NVIDIA

El script levanta postgres (pgvector) + ollama, crea la base codeindex con la extensión vector, descarga bge-m3, instala el server como uv tool y verifica cada punto al final (falla con mensaje claro si algo quedó a medias).

Después, una sola vez, el índice inicial (~30 min para ~7K archivos):

uv run python -c "from code_index_pg.indexer_pipeline import index_project; print(index_project('/ruta/a/tu/codigo'))"

Registro en Hermes — el wrapper espera a que postgres acepte conexiones antes de lanzar el server (sin eso, el server muere si la BD no está lista):

# Linux / WSL
hermes mcp add codeindex --command /ruta/al/repo/deploy/mcp-wrapper.sh

# Windows
hermes mcp add codeindex --command powershell --args -NoProfile -File C:\ruta\al\repo\deploy\mcp-wrapper.ps1

Solo los servicios van en Docker: el server MCP es stdio y corre en el host, usando los defaults del código (127.0.0.1:5432/codeindex, http://localhost:11434).

Run

# MCP server (stdio — used by the proxy wrapper)
code-index-pg --project-path ~/develop

# Full index (one-time, ~10 min for ~7K files)
python -m code_index_pg.indexer_pipeline  # via test harness, or:
uv run python -c "from code_index_pg.indexer_pipeline import index_project; print(index_project('/home/jam/develop'))"

MCP tools

Tool	Description
`set_project_path`	Index a project and start the incremental watcher
`update_index`	Incremental re-index: add missing files, drop deleted ones (no re-embed)
`search_code`	Literal/regex search over symbol names + file paths (paginated)
`search_semantic`	Embedding similarity search (bge-m3, pgvector cosine)
`callers_of`	Symbols that call the given symbol
`callees_of`	Symbols called by the given symbol
`impact_analysis`	2-level recursive callers + affected files
`find_files`	Glob search over indexed files
`get_symbol_body`	Signature, docstring, location of a symbol
`get_file_summary`	Language, line count, symbols of a file
`get_status`	Indexed projects + counts + active project
`remove_project`	Remove an indexed project and all its data in cascade (rejects the active project)

Hermes integration

Proxy: port 3104 in ~/.hermes/scripts/mcp-proxies.sh (supervisor systemd)
Wrapper: ~/.hermes/scripts/mcp-codeindex-wrapper.sh
Register: hermes mcp add codeindex --url http://127.0.0.1:3104/mcp

Environment

Var	Default	Notes
`CODEINDEX_DATABASE_URL`	`postgres://postgres:postgres@127.0.0.1:5432/codeindex`	local tool
`OLLAMA_URL`	`http://localhost:11434`	embeddings
`CODEINDEX_EMBED_MODEL`	`bge-m3`	must match pgvector dims (1024)
`CODEINDEX_PROJECT`	unset	active project (set via `set_project_path`)

Tests

uv run pytest tests/ -q   # unit + contract (in-memory MCP client)

Spec / OpenSpec

Living spec: openspec/specs/code-index/spec.md — archived changes in openspec/changes/archive/ (unified-code-index, project-directory-management).

License

MIT — free to use, modify and redistribute. See LICENSE.