Qartez MCP - codebase intelligence for AI agents

Most code-intelligence tools for agents are wrappers around grep. Qartez takes the opposite approach - it pre-computes a knowledge graph of your repo and hands the agent structured queries against that graph, so the agent stops re-reading the same files to figure out who calls what.

The pitch in the README is roughly a 94% reduction in tokens spent on code exploration versus naive file scans. That number is going to vary wildly by codebase, but the qualitative shift is real: PageRank and blast-radius signals let an agent ask "what's the most central thing in this module?" without opening a single file.

What it actually computes

The index is a single SQLite database at .qartez/index.db. Per file and per symbol, it stores:

• PageRank importance - symbols weighted by how many other symbols depend on them
• Blast radius - transitive count of symbols affected by a change
• Git co-change - which files historically move together (default: 300 commits of history, configurable via --git-depth)
• Cyclomatic complexity - per-function, for the 21 imperative languages it parses for hotspots

Tree-sitter parses 37 languages including TypeScript, Rust, Go, Python, Java, Kotlin, Swift, C/C++, Ruby, PHP, Elixir, Zig, plus config-shaped things like Terraform/HCL, Dockerfile, Helm, and systemd units. A file watcher keeps the index live by default; pass --no-watch to disable.

The modification guard is the part to pay attention to

Qartez ships three binaries: the MCP server, a setup wizard, and qartez-guard. The guard is the bit nobody talks about and it is the most opinionated thing in the project.

When an agent attempts an Edit, Write, or MultiEdit on a file whose PageRank exceeds 0.05 or whose blast radius is >= 10, the edit is blocked. The only way through is for the agent to first call qartez_impact file_path=<path>, which surfaces what would break. That call grants a 10-minute edit window for that specific file, after which the gate closes again.

You can tune or kill it via env vars:

QARTEZ_GUARD_PAGERANK_MIN=0.05    # raise to be less strict
QARTEZ_GUARD_BLAST_MIN=10
QARTEZ_GUARD_ACK_TTL_SECS=600     # the 10-minute window
QARTEZ_GUARD_DISABLE=1            # off

This is the most useful piece for anyone running agents unattended. The default thresholds force a "look before you leap" check on the load-bearing files - the ones you actually care about - while leaving everything else untouched.

Quick start

One-liner install (macOS, Linux, Windows via WSL):

curl -sSfL https://qartez.dev/install | sh

Or via Cargo:

cargo install qartez-mcp
qartez-setup

The setup wizard auto-detects 19 supported clients in one pass - Claude Code, Claude Desktop, Cursor, Windsurf, Zed, Continue.dev, and the rest. To target a subset:

qartez-setup --ide cursor,zed,claude

If you'd rather wire it up by hand, the Claude Desktop config looks like:

{
  "mcpServers": {
    "qartez": {
      "command": "/absolute/path/to/qartez",
      "args": []
    }
  }
}

When to reach for it

• You're running long-horizon agents that re-explore the same repo on every session.
• You have a few load-bearing modules that get edited too casually - the guard pays for itself the first time it blocks a sloppy edit.
• You want git co-change signals (which files move together) without paying the cost of running git log queries on every turn.

When not to

• Tiny repos. The guard and the index are both overkill if the agent can hold the whole codebase in its head.
• Languages outside the 37-language list. Tree-sitter coverage is broad but not universal - niche DSLs fall back to nothing.
• Workflows where you specifically want the agent to read source files (e.g. learning the codebase, doing a security review). The graph is a supplement, not a replacement.

Trade-offs and rough edges

Semantic search (the qartez_semantic tool) is gated behind an optional Cargo feature and pulls a ~270 MB model on first use. If you don't need vector search, skip it - most of the value is in the structural signals.

Re-indexing on a large monorepo isn't free. The watcher amortises this for you, but the first index build on a fresh clone is the kind of thing you kick off and walk away from.

There's no cloud component. Everything is local: SQLite for state, tree-sitter for parsing, no API key required. That's a feature, not an oversight.