
livespec
by Rixmerz
livespec
Public beta (v0.30). Local-first code intelligence for AI agents. Call graph, impact analysis, and Spec ↔ code traceability. Search is FTS5-only (no embeddings). Dead/legacy findings are graph evidence, not production traffic — confirm with APM/logs before deleting. Spec and docs mutation tools are plugin-gated. Pin:
uvx [email protected].License: GNU AGPL v3 (
AGPL-3.0-only). If you modify livespec and offer it over a network (MCP host, SaaS, internal API), you must provide the corresponding source under the same license. Corporate adoption notes and a first-contact compliance template:docs/AGPL_COMPLIANCE_CONTACT.md.
Code intelligence for AI agents — call graph, impact analysis, and bidirectional Spec ↔ code traceability (functional requirements, ADRs, NFRs, and other spec kinds). Local-first, zero external services.
<p align="center"> <img src="docs/assets/spec-explorer.png" alt="livespec Spec Explorer — Spec list, coverage, and linked symbols" width="920" /> </p> <p align="center"><em>Spec Explorer (docs plugin) on this repo — Specs, coverage, and linked symbols at a glance.</em></p>Who this is for / not for
| For | Not for |
|---|---|
| Agents cold-opening an unfamiliar repo | “Delete all unused code” without APM/logs |
| Impact analysis before a refactor or PR | Deep semantic search (vectors removed in 0.29) |
| Spec ↔ code / OpenSpec traceability | Indexing a parent folder of many unrelated repos |
Polyrepo HTTP flows via group_db | Replacing tests, human review, or runtime debugging |
Ships as a Claude Code plugin bundling three things: the MCP server
(the tools), a specialized subagent, and a preloaded Skill (the
operating manual). Everything an agent sees — the plugin, subagent, Skill, and
MCP tool namespace — is livespec. The PyPI distribution and console command
are also livespec (uvx [email protected] / pip install livespec).
The legacy command alias livespec-mcp still works on the same entry point.
Battle-tested on real codebases. Four releases of compounding wins on the same Django 5.1.4 queries:
| Tool on Django (40K symbols) | v0.8 | v0.9 | v0.10 | v0.14 |
|---|---|---|---|---|
find_dead_code candidates | 824 | 514 | 348 | 344 (−58% cumulative) |
find_endpoints(framework='django') | 20 | 162 (+8×) | 162 | 162 |
quick_orient p95 | ~60 ms | ~60 ms | ~60 ms | ~60 ms |
| Partial reindex (touch 1 file, Django) | — | ~7 s | — | 1.4 s |
Validated across 5 distinct agent profiles (exploration, refactor,
Spec flow, Django bugfix, TypeScript feature) — see
docs/AGENT_USAGE_DATA.md.
Want the agentic flow without reading further?
- MCP prompt
agent_playbook— how to use tools and annotate code (@spec:).docs/AGENT_QUICKSTART.md— 10-step cold-open flow.docs/AGENT_PLAYBOOK.md— same content as the prompt.
30-second tour
# Wire as an MCP server next to your editor (claude.ai/code, Cursor, ...).
# Every tool call carries workspace="/abs/path" — one server, N repos.
uvx [email protected]
# alias still works: livespec-mcp
// First call: cold-index the workspace once
> index_project()
{ "files_total": 2898, "symbols_total": 39789, "edges_total": 465179,
"languages": {"python": 2786, "javascript": 112}, "watcher_started": false }
// Composite first-contact on an unfamiliar symbol
> quick_orient(qna
Related servers

n8n
by n8n-io
Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.

@modelcontextprotocol/server-everything
OfficialMCP server that exercises all the features of the MCP protocol

@modelcontextprotocol/server-filesystem
OfficialMCP server for filesystem access