xmp4 — Semantic code knowledge for your stack

The 30-second pitch

Your AI coding agent is burning tokens grepping OSS libraries it will barely use. xmp4 is a hosted MCP server that pre-indexes 856 popular open-source libraries with SCIP — the semantic code format Sourcegraph uses — and serves them through 17 tools. No clone. No grep. No false positives.

ASK: "Who calls Flask.wsgi_app in the flask repo and what does it do?"
with grep + local clone:  git clone flask/flask          ~40 MB,   ~2 min  grep -rn "wsgi_app" .          200+ matches, mostly noise  cat src/flask/app.py | sed ... read 1000+ lines to find the body  filter false positives         model spends tokens deciding what's real  ──────────────────────────────────────  total:                         ~15,000 tokens + disk + wall time
with xmp4:  xmp4_info(symbol_name="Flask",    file_path="src/flask/app.py")         → signature,  20 tok  xmp4_source(symbol_name="wsgi_app", file_path="src/flask/app.py")       → body,      180 tok  xmp4_callers(symbol_name="wsgi_app", file_path="src/flask/app.py")      → 1 caller,   50 tok  ──────────────────────────────────────  total:                                                                    ~250 tokens
xmp4 is 60× cheaper here — and every result is SCIP-resolved, not text-matched.

The measured numbers (4 big OSS libs · reproducible)

Same realistic question on spring-boot · tokio · django · efcore: "give me the signature, body, and real callers of X."

GitMCP and Context7 look cheaper per call because they return less. To reach the same answer, GitMCP balloons to 42× more tokens — and still can't produce the semantic caller list. Context7 can't at any cost. Full whitepaper with Python harness →

Connect in 30 seconds

// Claude Code / Cursor / Claude Desktop — project `.mcp.json` or client config{  "mcpServers": {    "xmp4": {      "type": "http",      "url": "https://mcp.example4.ai/mcp"    }  }}

(The ready-to-paste config also lives at .mcp.json in this repo.)

Teach Claude how to use xmp4 (optional but recommended)

Install the xmp4 skill once per version — Claude will pick the cheapest tool path automatically (tests_for + view over grep):

# Claude Codemkdir -p ~/.claude/skills/xmp4 && \  curl -sfL https://example4.ai/xmp4-skill.md -o ~/.claude/skills/xmp4/SKILL.md
# Other clients: just tell Claude to read the URL when using xmp4 tools#   https://example4.ai/xmp4-skill.md

Restart your client. Then try (every step verified live 2026-04-24):

"Using xmp4, find the Flask class in flask/Flask and list its usages."

You should see Type Flask src/flask/app.py:81 and 165 usages across 33 result pages. One semantic call per question. Zero grep loops.

Setup for Cursor · Claude Desktop · Continue · Windsurf → docs/connect-instructions.md

The 17 tools

Full reference with live examples

Semantic core (where the value lives) xmp4_projects · xmp4_search · xmp4_info · xmp4_usages · xmp4_callers · xmp4_callees · xmp4_hierarchy · xmp4_outline · xmp4_source · xmp4_tests_for · xmp4_deps · xmp4_symbol_at

Convenience xmp4_view (raw file excerpt by line range) · xmp4_grep (server-side regex when semantics isn't enough)

Meta xmp4_guide (returns a versioned skill pointer to https://example4.ai/xmp4-skill.md — fetch once per version and save as a local Claude Code skill; embeds a minimal cheatsheet as offline fallback) · xmp4_server (version + stats)

Language coverage

Tier 1 — full coverage contract
C# · TypeScript · Python · Java · Rust · PHP

Tier 2 — best-effort, documented quirks
Go · JavaScript · Dart · Ruby · C++

Every known limitation — empty hierarchy.base on TS/Rust/Java/PHP, Python cross-module usages under-count, C# explicit-interface-impl dotted-name behaviour — is listed verbatim in docs/tiers-and-quirks.md. We'd rather set expectations correctly than have a reviewer find a gap and assume the whole thing is inflated.

Coverage grows by demand, not by guesswork

The index currently holds 856 repositories / 15 921 SCIP-indexed projects. We add new libraries based on two signals, combined:

1. Aggregate query logs — symbol names and project filters, no PII, no user code. If many AI agents search for a library we don't have, we see it.
2. Your request — file a repo-request issue with the GitHub URL, the language, and one concrete query you want to run. A single user request + downstream query demand almost always means indexed within days.

A public /stats/top-missing endpoint is planned — full transparency on what drives the growth loop.

Privacy — short version

• ✓ We log: aggregate query counts (symbol/project/tool names, coarse timestamps) to grow the index by demand.
• ✗ We don't log: the contents of your codebase · personal identifiers · request bodies beyond declared tool parameters.
• Standard nginx access logs kept 7 days for abuse prevention, then purged. Not joined with query tallies.

Full detail → docs/privacy.md.

Status

• 🟢 Live — mcp.example4.ai v1.2.1 · 856 repos · 15 921 projects · 17 tools · 11 languages
• 🟢 Benchmark published — reproducible whitepaper with Python harness
• 🟢 Listed on the Official MCP Registry as ai.example4/xmp4 (DNS-authed on example4.ai)
• 🟢 Smithery — smithery.ai/servers/0ics-srl/xmp4
• 🟢 Cursor Directory — cursor.directory/plugins/lsai-xmp4public
• 🔧 More registry submissions in flight — MCP.so, mcpservers.org, [email protected], awesome-lists (PRs open on punkpeye/jaw9c/appcypher)
• 🔧 Demand-driven growth loop — in progress

What's in this repository

Related

• LSAI protocol — open spec for semantic code intelligence in AI agents.
• SCIP — the semantic code format xmp4 is built on (Sourcegraph-developed, BSD-3).
• Model Context Protocol — the open transport spec xmp4 speaks.

License

Apache 2.0 for this documentation repository — see LICENSE. The hosted mcp.example4.ai endpoint is free for personal and commercial use (TOS link pending).

Commercial licensing or self-hosted deployment enquiries → open a GitHub issue labelled commercial on this repo.