Flag mixed output-format instructions
Surface references to multiple output formats that may lead to hybrid output.
A linter for agent instructions
Catch ambiguous agent instructions before a model runs.
LintLang checks tool descriptions, system prompts, AGENTS.md, YAML, JSON, and supported Python prompt sources for bounded language and contract defects. Local, deterministic, zero-LLM.
Illustrative rendering of an exercised LintLang 0.4.1 scan
1
Keep trying until it works.2
Respond in JSON and Markdown.Clear instructions✦Better review✦Fewer surprises✦Clear instructions✦Better review✦Fewer surprises✦
THE INSTRUCTION GAP
Natural language is a beautiful interface. It is also a very loose one.
Agent configurations are becoming application code, but they are still written and reviewed like prose. LintLang gives your instructions the same disciplined feedback loop as the rest of your stack.
Explore the approach →Expose bounded risks
Spot vague tool descriptions and qualifiers, missing operational limits, and context-boundary risks.
Set a baseline
Bring repeatable checks to every change, from a local prompt edit to a CI pull request.
A TIGHTER FEEDBACK LOOP
Write. Check. Ship with intent.
- 1
Point at supported instructions
Scan recognized JSON/YAML agent-config shapes, .txt, .md, and .prompt files, or Python files with extractable prompts.
- 2
Read findings in context
Each finding reports a detected pattern, severity, logical location, and suggested review action. Findings may remain at file or logical-location level.
- 3
Keep the standard in CI
Turn configuration review from a heroic habit into a dependable part of your delivery flow.
START IN A MINUTE
Put a check between the draft and the deploy.
Run LintLang against the instruction file in front of you, then bring the same check into CI.
Working with a coding agent? Tell it: “Run LintLang against our agent instructions before changing them.”
Run once, no install · LintLang 0.6.0 · Python 3.10+
$ uvx --from lintlang==0.6.0 lintlang scan AGENTS.mdWHERE TO USE IT
Run the check where instructions change.
Each entry point ships in the LintLang repository and is pinned to the 0.6.0 release. Replace AGENTS.md with the instruction paths your repository owns.
GitHub Action
Scan a chosen path on every push or pull request, or run lintlang init --github to write a pinned Code Scanning workflow.
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
persist-credentials: false
- uses: hermes-labs-ai/lintlang@v0.6.0
with:
path: AGENTS.mdpre-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: v0.6.0
hooks:
- id: lintlang
args: [AGENTS.md]Claude Code plugin
A non-blocking PostToolUse hook returns repair guidance after Write or Edit changes a supported file.
$ pipx install lintlang==0.6.0
# then, inside Claude Code
/plugin marketplace add hermes-labs-ai/lintlang
/plugin install lintlang@lintlangGemini CLI extension
A non-blocking AfterTool hook returns repair guidance after write_file or replace changes a supported file. Requires uv; the guide carries the commit-pinned install command.
Extension guide ↗MegaLinter
Opt-in external plugin AI_LINTLANG for repositories that already run MegaLinter.
GitHub Code Scanning
SARIF 2.1.0 output places instruction findings beside code findings.
$ lintlang scan AGENTS.md --format sarif \
--fail-on fail > lintlang.sarifThe 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.
IN THE FIELD
Check the language before runtime.
LintLang brings static analysis to the language agents execute: before runtime review and evaluation.
Public Larch CI
Character.AI's public Larch repository added LintLang to its consolidated third-party CI lint job: 236 files passed with no HIGH or CRITICAL findings.
Read the merged PR ↗8,508 downloads
All-time PyPI downloads including mirrors; 4,269 downloads excluding known mirrors. Coverage through August 21, 2026.
Inspect the PyPI stats ↗Gentoo packaging
Packaged in the community-maintained Haven overlay for Gentoo, with ebuilds through version 0.4.1.
Browse the ebuilds ↗Built in public
An Apache-2.0 package with public source, designed to fit a local-first engineering workflow.
Explore the repository ↗Evidence, not endorsements:Public receipts checked August 22, 2026; PyPI all-time counts cover through August 21.
Directory listings, not integrations or endorsements:MegaLinter External Plugins Catalog ↗·awesome-gemini-cli ↗·Product Hunt ↗
A DIFFERENT KIND OF LINTER
Built for the messy parts of language.
LintLang is designed around a simple idea: instructions deserve clear, actionable feedback without pretending that language is perfectly deterministic.
Inspect evidence & provenance ↗01
Human-readable findings
Feedback for the people who write and maintain agent behavior.
02
Severity with judgment
Focus attention where an instruction failure can carry real consequences.
03
Local-first workflow
Check the configuration in front of you, then bring it into your pipeline.
THE BOUNDARY
Static analysis narrows risk. It does not certify an agent.
A clean scan does not prove runtime correctness or safety. Use LintLang before evaluations, traces, guardrails, and human review—not instead of them.
See where it fits in the stack ↗
START WITH THE WORDS