Saurix

Joining an unfamiliar repo and working out what a change breaks means opening file after file. Saurix indexes the repo into a knowledge graph you can query, so the answer comes from one lookup instead of fifty file reads.

It plugs into Claude Desktop, Cursor, and OpenCode through the Model Context Protocol. saurix-mcp serves the graph and your agent asks the questions.

See it run

CLI	Dashboard	Agent tools

One number to hold or doubt: on pallets/click at 874ca2bc, 7.35% of the calls its own test suite actually made were edges Saurix had inferred. That is precision on observed calls, not recall, and every miss is listed in the open. Method, limits, and the second figure from a hand-labeled sample: docs/accuracy.md.

To see the whole loop an agent follows, read docs/agent-lifecycle.md. Every call there ran against a real repo. For why the project exists and where it falls short, read docs/why-saurix.md.

1. What it does

• Indexing turns a local path or a GitHub URL into a Graph of Symbols and Edges. Four Extractors (Python, TypeScript, Go, Java) all parse through tree-sitter.
• Queries answer structural questions without reading every file. Find where a Symbol is defined, who calls it, what it reaches, how two Symbols connect, and what a change would touch.
• Impact is the reverse reachable set from a Symbol through CALLS and CONTAINS edges. It reads as a blast radius: the files and functions to re-read after an edit.
• MCP tools expose all of it to agents, with one response shape everywhere. See section 5.

2. Architecture

Indexing takes the checkout, sends each file to its language Extractor, and merges the results into a Graph persisted as JSON. Queries, Impact analysis, and the 2D dashboard all read that Graph. Full breakdown: docs/architecture.md.

3. Setup

From PyPI:

pip install saurix
cd /path/to/your/project
saurix init

init indexes the project, writes the Graph to saurix.graph.json, generates the dashboard as saurix.html, and prints the MCP config for your client. From a source checkout, run uv sync first and prefix the commands with uv run.

To serve the Graph to an agent (Claude Desktop, Cursor, OpenCode):

saurix-mcp

The client starts and stops the server on its own. No manual terminal work needed beyond the config init printed.

4. Commands

Inside saurix, or as one-shots (saurix index ., saurix stats):

5. Agent tools (MCP)

The server exposes eight tools:

They answer three kinds of question:

1. Context discovery: find_symbol and related_files.
2. Behavioral mapping: callers, callees, and path_between.
3. Risk assessment: impact_of_symbol.

Every tool answers with the same envelope: ok, then either data or an error of code and message (never both), then meta with the duration_ms the call took. Tools that answer with rows add a count of them. A failure is timed like a success, so the shape does not change with the outcome.

Reading a graph, or looking something up in it, fails the same way whichever tool was asked:

Two more sets of codes come from the tool's own work rather than its arguments: index_repo reports SOURCE_NOT_FOUND, PERMISSION_DENIED, or INDEX_FAILED, and a query that fails underneath a tool it had already accepted reports STATS_FAILED, FIND_FAILED, CALLERS_FAILED, CALLEES_FAILED, PATH_FAILED, IMPACT_FAILED, or RELATED_FAILED.

A Query that matches nothing is not one of these: find_symbol returns an empty list, and so does a path_between whose two ends the graph holds but does not connect. A tool asked about a specific symbol is stricter, and reports SYMBOL_NOT_FOUND rather than guessing which one was meant. Pass an id from find_symbol if a name is not enough.

A runnable copy of every call, with real responses, lives in demo-mcp.md. It is generated from a live server run, so copy any of it and it works.

6. Development

Running tests

uv run pytest

Regenerating the documentation

The tool-call demo (demo-mcp.md) and the agent walkthrough (docs/agent-lifecycle.md) are generated from a real run of the MCP server against this repository, so every identifier in them is one a reader can copy. Refresh them after touching the code or the tool surface:

uv run scripts/generate_mcp_demo.py

--check reports drift without writing. tests/test_docs.py re-runs the generator on every test run and fails when the committed documents no longer match the current code, or when any document names a module or file path that no longer exists.

Checking call edge accuracy

How accurate the inferred call edges are is measured, not asserted. The accuracy report (docs/accuracy.md) pins a target repository to one commit, traces its own test suite at runtime, and reports how many of the calls that really happened the pipeline also inferred. A hand-labeled sample covers the region no test reaches, and is reported as a second figure rather than folded into the first. Regenerate it with:

uv run scripts/accuracy.py

It is a published snapshot, not a CI gate: it clones a foreign repository and builds an environment for it, which is too slow and fragile to run on every push. The report says so, and says what it would take to change that.

Regenerating the demo assets

The recording and the three screenshots at the top of this README are captured from real runs, not drawn. scripts/capture_demo.py copies this checkout to a scratch directory, drives init, stats, find and impact through a pseudo-terminal, photographs the dashboard the run generated, and replays the MCP tool calls the documentation generator makes:

uv run scripts/capture_demo.py

It needs Pillow and a Chromium from Playwright, both in the dev group, and a monospaced font on the machine. The pinned force-graph bundle is fetched once and cached under tmp/, so a rerun needs no network. --check reports captures that are missing or have outgrown their size budget without taking new ones. What each file is, and how the dashboard capture doubles as the portfolio card image, is written up in docs/assets/README.md.

Testing the MCP server

You can test the MCP integration without a full IDE using the MCP Inspector:

1. Install the Inspector: npm install -g @modelcontextprotocol/inspector
2. Run the Server: npx @modelcontextprotocol/inspector uv run saurix-mcp
3. Interact: Open http://localhost:5173, click Connect, and use the Call Tool tab.

For step-by-step setup (Claude Desktop, Cursor, OpenCode), simply run saurix init in your project folder.

Saurix is built for the era of autonomous coding.