crux

A local-first context engine for coding agents.

crux watches folders, keeps a live index of your code (plus docs, tickets, and web pages), and serves it over MCP so any agent — loop, Claude Code, Cursor, Windsurf, Zed — can ask precise questions about your codebase instead of grepping blindly.

Status: Phase 1. The live index works end-to-end and serves MCP.

Install

curl -fsSL https://raw.githubusercontent.com/notshekhar/crux/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/notshekhar/crux/main/install.ps1 | iex

No runtime needed — it's a single static binary. Language grammars are fetched on first crux init rather than bundled, so the download stays small.

crux init .                 # index a repo, print agent config
crux search "retry webhooks"
crux symbol RetryPolicy
crux status                 # what's indexed, how fresh
crux list                   # every workspace you've indexed
crux serve .                # browse the index in a browser
crux mcp .                  # what your agent runs
crux upgrade                # update in place

Related MCP server: code-index

Inspecting the index

crux serve opens a local inspector — loopback only, since the index holds your source:

• Graph — the import graph, as modules or individual files, force-directed. Hubs are highlighted; isolated files are counted rather than drawn.

• Search — the same three-arm search the agent gets, showing which arms matched.

• Files — per file: every symbol, and the chunk boundaries search actually uses.

• Symbols — definitions, callers, callees. Labelled heuristic until SCIP lands.

• Inspect — what is in the index, per-tier coverage, and storage by table.

Uninstall with CRUX_UNINSTALL=1 bash install.sh.

Where it keeps things

Everything lives under ~/.crux — nothing is written into the repos you index:

~/.crux/
  ├── config.json      settings
  ├── grammars/        tree-sitter parsers, fetched on first init
  └── index/
       ├── registry.json
       └── <project>-<hash>.db

crux list shows every index and its size; crux forget <path> deletes one.

Indexing many repos at once

Point it at a folder of projects and it indexes all of them into one searchable index, applying each repo's own .gitignore:

crux init ~/code            # 20 repos, 21,415 files, 106s, 588 MB
crux search "leader election heartbeat" ~/code

Measured on the loop repo (469 files, 3.9 MB of TypeScript): cold index 3.4 s at 138 files/s, 13,455 symbols, query p50 2.1 ms / p95 3.2 ms.

Built: the durable queue, leader election, tree-sitter extraction for TS/TSX/JS/Python/Go/Rust, content-addressed storage, FTS5 + trigram, three-arm fused search, the filesystem watcher, five MCP tools, and the CLI. Not yet: the code graph (refs, pack), embeddings, and connectors.

Why

Every "context for AI" tool today does the same thing: chunk the repo, embed it, run cosine similarity, dump twenty chunks into the prompt. That approach cannot answer the questions agents actually ask — "who calls this?", "what breaks if I change this signature?" — because those are graph queries with no similarity signal at all.

crux runs five retrievers (BM25, trigram, vector, symbol table, path fuzzy), fuses them, and then expands along the code graph. Embeddings are roughly 20% of the answer quality. The graph is the moat.

The second half of the thesis is freshness. A stale index is worse than no index, because the agent can't tell the difference. crux watches the filesystem, queues work durably in SQLite, and verifies every returned span against the file on disk before handing it to an agent.

Read the plan

Shape of it

  Claude Code      Cursor        loop          ← MCP hosts (stdio)
       └──────────────┼─────────────┘
                crux mcp × N                   ← thin readers
                      │
               .crux/index.db                  ← SQLite (WAL) + durable job queue
                      │
              one elected leader               ← watcher + parse worker + embed worker