What is Chisel?
Chisel answers three questions a coding agent has after it changes code: which tests should run, which files are risky, and who knows this code. It scans a repository's tracked files with its own extractor (no tree-sitter), links each test to the functions and classes it calls, mines git history for churn, co-change coupling and authorship, and keeps the result in one JSON snapshot under `.chisel/store`. Agents reach it over MCP on stdio, or through a JSON CLI.
It registers 29 MCP tools in six groups; only the seven core tools (analyze, update, diff_impact and friends) are always on, and `CHISEL_DISABLE` or a line in the project's AGENTS.md hides the rest. A Claude Code hook can pipe every test run into `chisel capture`, which logs failures as issues, ranks the code that changed since the test last passed, and closes the issue with the fixing commit. Chisel is MIT licensed, built from the Rust standard library with no crates, and released v0.23.0 on 11 November 2026.
How to install Chisel
chisel serve-mcp (release binary from GitHub, with CHISEL_PROJECT_DIR set)Is Chisel worth it?
Most code-graph MCP servers help an agent read less code. Chisel aims at the step after the edit: the agent calls diff_impact, gets a ranked list of the tests that call the changed functions and then the tests of their callers, two levels up by default, and runs those instead of the whole suite or a guess. The maintainer's benchmark on Django's package, about 1000 code files, puts a cold analysis at roughly 8 seconds and a warm update at about 69 milliseconds.
The failure log is the more unusual part. With the PostToolUse hook installed, each test run the agent makes (cargo test, pytest, Jest, Vitest, go test, node:test) is parsed for failures without spending model tokens, duplicates are folded together, and each issue lists suspect changes with a one-line reason. The next session starts with at most two lines naming open issues, which helps an agent that has forgotten what broke yesterday.
The catch: adoption is close to nil. The repository has 7 GitHub stars and one fork, and 115 of its 120 commits come from one author, so you are betting on that person's pace. It is fast-moving, with 30 releases since v0.5.4 in March 2026, and the GitHub description still advertises 15 MCP tools while the README lists 29. Matching is heuristic rather than type-checked, the README warns that its risk scores are heuristics, and above 2000 code files per-function churn is skipped. The optional HTTP API (`chisel serve`) binds to 127.0.0.1 by default and, in our reading of the source, has no authentication, so do not expose it beyond the machine.
Our take: Chisel is a sharp side tool rather than a code-intelligence platform. If your agent already has GitNexus or codebase-memory-mcp for navigation, Chisel's test selection and failure tracking fill a gap they leave; turn off its lock and history groups so the two do not overlap.
Where it earns its place
A backend developer with a slow pytest suite
Stop the agent running every test after each small edit
Call analyze once, then have the agent call diff_impact after each change and run only the ranked tests, finishing with a full run before merge. Free; the saving is the minutes of suite time per edit.
A solo maintainer using Claude Code across several sessions a day
Keep track of which tests the agent broke and what fixed them
Run `chisel hook-install --write` to add the capture and SessionStart hooks; failures become deduplicated issues that close automatically with the fixing commit when the test passes again.
A tech lead reviewing an agent's pull request
See which changed files are risky and who should review them
Run `chisel triage` and `who_reviews` on the branch for churn, coupling, untested code and current owners. The scores are heuristics, so treat them as a starting list, not a gate.
Before you commit
- Setup
- Download a release binary for Linux, macOS or Windows (or `cargo build --release`), put it on PATH with git installed, add `chisel serve-mcp` with CHISEL_PROJECT_DIR to .mcp.json, then call analyze once.
- Integrations
- Any stdio MCP client, including Claude Code, Cursor, Windsurf and Cline; a JSON CLI; a separate local HTTP API; Claude Code hooks for test capture; a bundled SKILL.md. Extracts code from 12 language families and recognises pytest, Jest, Go test, Rust tests, JUnit and others.
- Data and privacy
- Runs locally and stores its snapshot in `<project>/.chisel/store`; there is no hosted service. The HTTP API binds 127.0.0.1 by default with no authentication. Custom extractors named in a repo file run only when explicitly allowed with --allow-repo-extractor.
- Cost at scale
- Free under MIT, with no paid edition. Large repositories can be split into shards, each with its own snapshot.
- Limits
- Scores are heuristics; per-function churn is skipped above 2000 code files; automatic refresh runs inline only at or under 50 changed files; worktree answers are not supported with shards configured.
What’s good
- Ranks the tests an edit reaches, through callers two levels up
- Logs agent test failures as issues with suspects, no model tokens
- One dependency-free Rust binary; git is the only requirement
- Tool groups trim the 29 tools down to 7 core ones
What’s not
- 7 stars and a single main author
- Heuristic matching, not type resolution
- Per-function churn stops above 2000 code files
- HTTP API has no authentication
Best for
Developers who let a coding agent edit a well-tested repository and want it to run the right tests, not all of them.
Chisel alternatives
Chisel: common questions
- Is Chisel worth it?
- Chisel does one useful thing that bigger code graphs skip, picking the tests an edit actually touches, but it is a one-person project that very few teams have tried. We rate it 3.5 out of 5.
- How much does Chisel cost?
- Free and open source (MIT).
- What are the downsides of Chisel?
- 7 stars and a single main author; heuristic matching, not type resolution; per-function churn stops above 2000 code files; hTTP API has no authentication.
- Who is Chisel best for?
- Developers who let a coding agent edit a well-tested repository and want it to run the right tests, not all of them.
- How much does Chisel cost for a team or in production?
- Free under MIT, with no paid edition. Large repositories can be split into shards, each with its own snapshot.
- Is Chisel safe for company data?
- Runs locally and stores its snapshot in `<project>/.chisel/store`; there is no hosted service. The HTTP API binds 127.0.0.1 by default with no authentication. Custom extractors named in a repo file run only when explicitly allowed with --allow-repo-extractor.
- What are the best Chisel alternatives?
- The closest alternatives we have reviewed are code-review-graph, codebase-memory-mcp and GitNexus.
Answers checked Nov 13, 2026.