sextant

Sextant

Code-quality grader for AI-agent workflows. Deterministic rules and LLM judges, surfaced inside the agent loop and on every PR.

📖 Documentation →

What is Sextant, and when should I use it?

Sextant is a code-quality grader built for codebases where AI agents (Claude Code, Cursor, Copilot, …) write a real share of the code.

Imagine you want every diff — agent-written or human-written — to stay within your team's bar for file length, function length, complexity, duplication, test coverage, and whatever other house rules you have.

Usually you only find out a diff is over the line one of three ways:

• a teammate catches it in PR review, hours or days later;
• CI lint or format jobs flag it after the agent has already finished its turn and moved on;
• you read the diff yourself, decide the function is too long or duplicates something, and either fix it by hand or feed the complaint back to the agent as a follow-up prompt.

All three are reactive. By the time the feedback lands, the agent's context is cold and the cleanup falls on you.

Sextant moves that feedback into the loop. The same grader runs in two places:

• inside the agent, as an MCP tool and Claude Code hooks — the agent grades its own diff after every edit and self-corrects before ending the turn;
• on every PR, as a GitHub Action that posts a single review comment and only blocks merge on regressions, so pre-existing debt doesn't gate new work.

One .sextant/ config drives both. You customize:

• what counts as a finding — thresholds for the built-in size, complexity, duplication, and untested-public-function rules, plus your own markdown rules judged by Claude or GPT against individual files;
• how strict the verdict is — how many errors or warnings flip a grade from approve to request_changes;
• what the grader looks at — just the diff (default; fast, ignores legacy debt), the touched files, or the whole repo.

If you're hand-reviewing every agent diff for the same handful of issues, or watching CI flag the same things over and over, Sextant is the loop you're missing.

One engine, five surfaces

• MCP server — agents call grade_diff after each edit and self-correct before finishing the turn.
• GitHub Action — posts a single review comment on every PR. Gates merge on regression — pre-existing findings don't block new work.
• Claude Code plugin — skills, slash commands, and hooks that wire grading into the edit loop with zero extra configuration.
• Editor (LSP) — sextant-lsp powers a VS Code extension (and any other LSP client) with live squiggles and hover-to-explain showing the full rule body.
• CLI — sextant grade for human, JSON, markdown, or SARIF output. Works offline once installed.

Highlights

• Built for the inner loop. Diff-mode grading runs in well under a second on a typical change, so an agent can grade after every edit without slowing the user down.
• Deterministic by default. Seven built-in rules cover file length, function length, parameter count, cyclomatic complexity, nesting, token duplication, and untested public functions — pure Rust, no network calls.
• LLM rules when you want them. Author markdown rules evaluated by Claude or GPT against individual files. Cached by content hash so repeat grades are free.
• Regression-aware. PR mode only blocks on findings introduced by the change. The repo's existing debt is its problem; the PR's job is to not make it worse.
• Multi-language. Tree-sitter parsers for Rust, Python, Go, Java, TypeScript, TSX, and JavaScript.

Quickstart

Install the binaries (release archives, Homebrew tap pending, or from source):

cargo install --path crates/sextant-cli
cargo install --path crates/sextant-mcp
cargo install --path crates/sextant-lsp   # editor integration

Bootstrap a config and grade your first repo:

cd your-repo
sextant init
sextant grade                                 # whole-file mode
sextant grade --diff --base origin/main       # only changed lines
sextant rules list                            # what rules are loaded

For full installation options (releases, brew, source) see Installation.

Use it inside Claude Code

/plugin marketplace add kylebastien/sextant-mcp
/plugin install sextant@kylebastien/sextant-mcp

The plugin registers the MCP server, three skills, three slash commands, and three hooks that grade after every edit and before the agent ends its turn. Read the plugin guide for the details.

Use it on every PR

# .github/workflows/sextant.yml
name: Sextant
on:
  pull_request:

permissions:
  contents: read
  pull-requests: write