Skip to content
LintLang 0.8.3 · Python 3.10+ · Apache-2.0Local · deterministic · zero LLM

Static analysis for agent language

Lint the instructions your agents execute.

LintLang flags ambiguous tools, missing bounds, schema gaps and mixed output formats in AGENTS.md, prompts, tool definitions, configuration and supported Python prompt code—before a model runs.

Run onceno install
$ uvx --from lintlang==0.8.3 lintlang scan AGENTS.md
See findings for the browser example ↓
Inputs
Text · YAML · JSON · Python
Files, directories, or stdin
Outcomes
PASS · REVIEW · FAIL · ERROR
Nonblocking by default
CI formats
Terminal · JSON · SARIF 2.1.0
Stable rule IDs and honest locations
Runtime
PyYAML only
No model or API calls; input stays local

This embedded browser demo runs the verified LintLang 0.6.0 wheel. The full playground and install commands use the current 0.8.3 release. The first browser scan downloads Python and the engine; scan input stays in your tab.

Browser demo · LintLang 0.6.0

Put one real instruction set through the linter.

Run the example below, or paste your own prompt, tool definition, or configuration. The actual LintLang engine runs locally in your browser. For the current release and five editable experiments, open the full LintLang playground ↗.

Local executionZero LLM calls
01 · Input

Your instructions

437 / 20,000
Example reference · expected resultFAILH1.2 · H2 · H1.6Compare the original and improved examples by running a scan on each.
Input stays in this browser and is never stored.

The first scan downloads the Python runtime and LintLang, so it may take a moment. Later scans reuse the loaded engine.

02 · Result

Deterministic findings

Ready when you are.

Run the loaded example to see its findings. Then load the improved example and scan again to compare.

No model · no upload · no account

Recorded result for the default example. The LintLang 0.6.0 wheel that runs this demo returns FAIL for the original two-tool example:

  • [HIGH] H1.2 tool:search_guides — Tool 'search_guides' has a very short description (18 chars): "Search the guides."
  • [HIGH] H2 system_prompt — System prompt defines tools but contains no termination conditions, retry budgets, or progress checks.
  • [MEDIUM] H1.6 tool:search_docs vs tool:search_guides — Tools 'search_docs' and 'search_guides' carry no differentia — every meaning-bearing term in one is present, or has a synonym, in the other. Both descriptions may be accurate and still give a model nothing to choose between them.

The improved example returns PASS with no findings. These are structural findings about the text; they do not show how a model would select tools or when it would stop at runtime.

01

Ambiguous tools

Descriptions that are empty, vague, overlapping, or missing the words that distinguish neighboring tools.

02

Missing constraints

Retry loops, tool-using instructions, and workflows that omit limits, stop conditions, or progress checks.

03

Contract mismatches

Tool descriptions that disagree with schemas, conflicting output requirements, and malformed message roles.

04

Embedded prompts

Prompt text and uncalibrated thresholds extracted from Python source without executing the application.

1Run onceuvx --from lintlang==0.8.3 lintlang scan AGENTS.md
2Review the baseline

Fix findings or record why the language is intentional.

3Gate deliberatelylintlang scan AGENTS.md --fail-on fail

To check every recognized instruction file in a repository instead of one path, run uvx --from lintlang==0.8.3 lintlang scan --discover .. Discovery finds AGENTS.md, CLAUDE.md, GEMINI.md, SKILL.md, agent.yaml/.yml/.json, GitHub Copilot instruction files, and *.mdc project rules under .cursor/rules/. Names and documented layout segments are case-sensitive; other editor layouts need an explicit path. What gets scanned ↗

A normal scan reports findings without failing the process. --fail-on failblocks only HIGH and CRITICAL findings; --fail-on review also blocks MEDIUM findings. Missing or malformed requested inputs remain errors either way.

GitHub Action

Scan a chosen instruction path on every push or pull request. Since 0.8.2 the Action is advisory by default; add fail-on: fail to block on HIGH or CRITICAL findings. lintlang init --github writes a pinned Code Scanning workflow that gates by default instead.

- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
  with:
    persist-credentials: false
- uses: hermes-labs-ai/lintlang@v0.8.3
  with:
    path: AGENTS.md
Action setup ↗

pre-commit

Scan the configured paths before each commit. Findings stay advisory until you add --fail-on fail.

repos:
  - repo: https://github.com/hermes-labs-ai/lintlang
    rev: v0.8.3
    hooks:
      - id: lintlang
        args: [AGENTS.md]
pre-commit setup ↗

Claude Code plugin

A non-blocking PostToolUse hook returns repair guidance after Write or Edit changes a supported file.

$ pipx install lintlang==0.8.3

# then, inside Claude Code
/plugin marketplace add hermes-labs-ai/lintlang
/plugin install lintlang@lintlang
Plugin guide ↗

Gemini CLI extension

A non-blocking AfterTool hook returns repair guidance after write_file or replace changes a supported file. Requires uv; the current guide remains pinned to the upstream-verified 0.5.3 source.

Extension guide ↗

MegaLinter

Opt-in external plugin AI_LINTLANG for repositories that already run MegaLinter.

MegaLinter setup ↓

GitHub Code Scanning

SARIF 2.1.0 output places instruction findings beside code findings.

SARIF setup ↓

Replace AGENTS.md with the instruction paths your repository owns. The editor hooks never block or rewrite a file. These integration points are maintained by Hermes Labs in the LintLang repository; they are not endorsements by GitHub, Anthropic, Google, or MegaLinter.

GitHub Code Scanning

Put instruction findings beside code findings.

Since 0.4.1, LintLang emits deterministic SARIF 2.1.0 with stable rule IDs, repository-relative paths, and evidence-supported line spans where the parser can justify them. Structured inputs remain file-level rather than inventing precision.

Copy the least-privilege workflow ↗
Generate SARIF0.8.3
$ lintlang scan AGENTS.md \
    --format sarif \
    --fail-on fail \
    > lintlang.sarif

MegaLinter · opt-in external plugin

Check agent instructions in your existing MegaLinter workflow.

Start with MegaLinter already installed and running in your repository. Merge these entries into .mega-linter.yml, preserving your existing plugins and linter selections. If you do not already use ENABLE_LINTERS, list the other linters you want to keep too: this setting limits which linters run.

AI_LINTLANG selects conventional instruction files such as AGENTS.md, CLAUDE.md, and prompt or tool files by name. The guide explains how to adjust that scope. It runs lintlang scan --fail-on fail: FAIL exits nonzero; REVIEW stays advisory.

Read the MegaLinter setup guide ↗
.mega-linter.ymlMerge with existing configuration
PLUGINS:
  - https://raw.githubusercontent.com/hermes-labs-ai/lintlang/main/mega-linter-plugin-lintlang/lintlang.megalinter-descriptor.yml

ENABLE_LINTERS:
  - AI_LINTLANG

Independent use and packaging

Recurring public CI

Character.AI's public Larch repository pins lintlang 0.3.1 as a blocking step in recurring CI. The linked public job records a successful execution, and Larch's linting reference links hermes-labs-ai/lintlang as the upstream.

Larch CI workflow ↗ · Historical Larch requirements pin ↗ · Successful public execution ↗ · Larch upstream linting reference ↗

Evidence class. Independent recurring CI execution

Unofficial Haven overlay

An independent maintainer provides Gentoo packaging in the unofficial Haven overlay, which tracks upstream releases across successive ebuild bumps. Its metadata.xml records hermes-labs-ai/lintlang as the upstream remote.

Haven overlay package ↗ · Upstream package metadata ↗

Evidence class. Independent packaging

Attributed product influence

The Agent Lint maintainer attributes parts of that product's roadmap to useful ideas from lintlang.

Agent Lint maintainer attribution ↗

Evidence class. Attributed product influence

Directory listings, not integrations or endorsements

Release and method

The boundary

Static analysis can narrow risk. It cannot certify an agent.

LintLang does not run models, evaluate runtime behavior, prove semantic correctness, or establish that an agent is safe in production.

Use it before prompt tests, guardrails, and production monitoring—not instead of them.

LintLang 0.8.3 is published and maintained by Hermes Labs. Canonical repository: https://github.com/hermes-labs-ai/lintlang.

See where static linting fits in the stack →Compare LintLang with adjacent tools →简体中文快速上手 →

Start with one real instruction file

Catch the defect while it is still just text.

$ uvx --from lintlang==0.8.3 lintlang scan AGENTS.md