hep-index

A multi-codebase code index for HEP software, served to AI agents over MCP.

HEP tools are notoriously hard to navigate — sparse docs, heterogeneous APIs, moving upstreams. hep-index lets an agent read the source instead of guessing: you register codebases (local checkouts, or fetched from a curated catalog of pinned upstream sources), and any MCP-capable harness (Claude Code, Codex, pi, Cursor, …) gets search_code / find_symbol / read_file tools over the whole stack at once.

It is deliberately not RAG: no embeddings, no vector store — ripgrep + ctags/sqlite symbol indexes over the real trees, live-read. Symbol-level lookup works across C++, Python, Fortran and (via search) Mathematica.

Install (consumers)

Zero-install, straight from the git remote (pin a tag for reproducibility — SSH key required for private repos):

# one-shot use
uvx --from git+ssh://git@github.com/BSMArt-HEP/hep-index.git@v0.3.0 hep-index list

# or install persistently (isolated venv, `hep-index` on PATH)
uv tool install git+ssh://git@github.com/BSMArt-HEP/hep-index.git@v0.3.0
hep-index list

Requirements: uv, Python ≥ 3.11, ripgrep (required for search), universal-ctags (optional — without it you get degraded manifest-only indexes). Updates: uv tool upgrade hep-index.

Related MCP server: Satori

Quickstart

# fetch + index a curated package (MadAnalysis 5, pinned upstream)
hep-index fetch madanalysis

# or index something already on disk, e.g. your project's MG5 install
hep-index add ~/recast/read/madgraph --name madgraph --version 3.7.0 --recipe madgraph

hep-index list          # what's indexed, staleness, git heads
hep-index catalog       # 29 curated sources w/ categories + gotcha notes

Serve it to an agent harness (stdio MCP):

hep-index serve

MCP wiring (any MCP-JSON client; adjust syntax to your harness). Preferred — zero-install via uvx, pinned tag:

{"mcpServers": {"hep-index": {
  "command": "uvx", "args": ["--from", "git+ssh://git@github.com/BSMArt-HEP/hep-index.git@v0.3.0",
                             "hep-index", "serve"]
}}}

Dev alternative (runs the live checkout): "command": "uv", "args": ["run", "--project", "/path/to/hep-index", "hep-index", "serve"].

Tools exposed

Wrong selector names self-correct: KeyError: not registered: 'madgrap' (available: …).

Example session (agent's view)

Q: does RecLeptonFormat have d0sig()? search_code(r"\bd0sig\b", codebase="madanalysis") → [] — no such method. find_symbol("d0error") → tools/SampleAnalyzer/Commons/DataFormat/RecLeptonFormat.h:202. read_file(that path, center=202, before=10, after=5) → the accessor + comment.

Every fact rediscovered from source in three calls, with file:line citations — no hand-maintained API notes to rot.

Layout & state

~/.config/hep-index/codebases.toml   # registry (machine-local, hand-editable)
~/.cache/hep-index/<name>-<version>/ # per-codebase sqlite index (rebuildable)
src/hep_index/data/catalog.toml      # curated source pins (shareable data)

Env overrides: HEP_INDEX_CONFIG, HEP_INDEX_CACHE, HEP_INDEX_CATALOG, HEP_INDEX_ROOT (fetch destination, default ~/hep-codebases).

Catalog = pins-as-data: canonical upstream URLs, known-good versions, and gotchas (dead tags, launchpad series-path quirks, Anubis walls) as commented TOML — contributions welcome via PR. Codebases themselves are never bundled or redistributed; each machine fetches from upstream.

Extending the catalog

The catalog is just commented TOML (src/hep_index/data/catalog.toml) — pins, not code. To propose a new package:

1. Add the entry (category from the existing set, versions with latest LAST — never a beta/pre-release):

[mytool]
category = "model-builder"          # me-generator | parton-shower | detector-sim |
                                    # analysis-framework | recast-engine | spectrum-calc |
                                    # model-builder | dark-matter | stat-interpretation |
                                    # validation | pdf | global-fit | symbolic-toolkit
type = "git"                         # or "tarball" (tar and zip both supported)
url = "https://github.com/<org>/mytool.git"
versions = ["v1.2.3"]
notes = "tag scheme vX.Y.Z; PyPI 'mytool' exists but source is canonical for navigation"

Extra fields when needed: recipe (preset exclude globs), extract_root (tarball top-level dir — leave unset to auto-detect; flat archives extract into the dest dir), urls (per-version URL overrides, e.g. launchpad series paths that aren't mechanical).

2. Verify before you PR — nothing enters unverified:

# git entries: exact tag string must exist
git ls-remote --tags --refs https://github.com/<org>/mytool.git

# tarball entries: direct URL must return 200 (Anubis-walled *pages* are fine,
# the file URLs usually still fetch)
curl -sIL -o /dev/null -w "%{http_code}\n" <tarball-url>

Hard-won lessons encoded in existing notes — read a few entries first: canonical upstream only (personal mirrors rot: restrepo/*, HEPcodes/*); claimed versions sometimes don't exist (SoftSUSY "4.1.24" didn't); dead release lines lurk (MadAnalysis v2.0.4_beta, frozen 2022); moving endpoints are allowed but must be flagged (FeynMaster's php latest stream, BSMArt pre-tag main).

3. Test the fetch in a scratch registry (never your real one):

HEP_INDEX_CONFIG=/tmp/c.toml HEP_INDEX_CACHE=/tmp/c HEP_INDEX_ROOT=/tmp/r \
  uv run hep-index fetch mytool

4. PR checklist: entry parses (uv run hep-index catalog), fetch works in scratch, version is latest-LAST stable, notes name the hosting gotcha, and uv run pytest stays green (bump the entry-count assertion in tests/test_catalog.py).

Development

git clone git@github.com:BSMArt-HEP/hep-index.git && cd hep-index
uv sync
uv run pytest        # 91 tests, offline (file:// fixtures)
uv run hep-index …   # CLI from the live checkout

Architecture and binding interface contracts live in CONTRACT.md; the full HEP-software landscape map (incl. Tier-2/3 packages not yet catalogued) in docs/HEP_SOFTWARE_MAP.md.

AI-assistance disclaimer

This codebase was developed with substantial AI assistance: implementation, tests, and the HEP-software landscape research were produced by LLM agents (GLM models, orchestrated via the pi coding agent) working from human-authored interface contracts, with human review of every change. Upstream packages referenced in the catalog are the work of their respective authors and are only pinned, never redistributed.

License

MIT — Copyright (c) 2026 Miguel Crispim Romão. Catalog entries pin upstream software by other authors; their licenses apply to their code, which is never redistributed here.