Skip to content

Get started

Pick the way you use agents. For the coding-agent plugins you need Python 3.10+ and git on the machine; the plugin builds its own environment the first time it starts (about 10 seconds). Setting it up for a team? Read operating it too.

Install it by hand:

claude plugin marketplace add https://github.com/AgentLTL/agentltl-claude-code.git
claude plugin install agentltl@agentltl

Then, in Claude Code, run /reload-plugins and /agentltl:setup.

Or let Claude do it: paste this into Claude Code. It installs the plugin, turns on updates, and helps you choose the first rules. (It follows the setup file at that address, so read it first if you want to know what it will do.)

Set up the AgentLTL plugin for me by following https://raw.githubusercontent.com/AgentLTL/agentltl-claude-code/main/SETUP.md

The plugin does nothing until there is an AGENTLTL.yaml, at a project's root for that project, or in ~/.claude/ for every project. Setup writes the first one with you.

Check that it is on: the status line shows AgentLTL ● 4 rules · … once you run agentltl statusline --install (setup offers it), and /agentltl:status lists the rules in force.

More on the Claude Code plugin

Install the plugin from its marketplace:

copilot plugin marketplace add AgentLTL/agentltl-copilot-cli
copilot plugin install agentltl@agentltl

Then, in a Copilot CLI session, run /agentltl-setup to choose the first rules.

The plugin does nothing until there is an AGENTLTL.yaml, at a project's root for that project, or in ~/.copilot/ for every project. Setup writes the first one with you. A project's file works for Claude Code, Copilot CLI, Mistral Vibe and Codex alike.

Check that it is on: run agentltl validate in the project. It lists the rules in force and where each comes from.

More on the Copilot CLI plugin

Clone the plugin into Vibe's plugin folder, then add its hooks:

git clone --recurse-submodules https://github.com/AgentLTL/agentltl-mistral-vibe.git ~/.vibe/plugins/agentltl
~/.vibe/plugins/agentltl/bin/agentltl install

install adds AgentLTL's hooks to ~/.vibe/hooks.toml (your own are kept) and links agentltl into ~/.local/bin. Then, in a Vibe session, run /agentltl:setup to choose the first rules. It works with any model Vibe runs, Mistral's or an OpenAI-compatible one of your own.

The plugin does nothing until there is an AGENTLTL.yaml, at a project's root for that project, or in ~/.vibe/ for every project. Setup writes the first one with you.

Check that it is on: run agentltl validate in the project. It lists the rules in force and where each comes from.

More on the Mistral Vibe plugin

Install the plugin from its marketplace, then trust its hooks:

codex plugin marketplace add AgentLTL/agentltl-codex
codex plugin add agentltl@agentltl
"$(ls -td ~/.codex/plugins/cache/agentltl/agentltl/*/ | head -1)"bin/agentltl install

Codex runs no hook until you trust it: install trusts AgentLTL's hooks in ~/.codex/config.toml (the rest of the file is kept) and puts agentltl in ~/.local/bin. You can approve each hook in Codex's /hooks instead. Then, in a Codex session, run $agentltl:setup to choose the first rules. It works with any model Codex runs, OpenAI's or your own.

The plugin does nothing until there is an AGENTLTL.yaml, at a project's root for that project, or in ~/.codex/ for every project. Setup writes the first one with you.

Check that it is on: run agentltl validate in the project. It lists the rules in force and where each comes from.

More on the Codex plugin

pip install agentltl                 # the formula engine, no dependencies
pip install "agentltl[smolagents]"   # with run-time enforcement for smolagents
pip install "agentltl[native]"       # or the native OpenAI-compatible loop

from agentltl import Constraint, Before, CalledNTimes, verify_trace

trace = {"tool_calls": [{"tool_name": "fetch_data"}, {"tool_name": "save_results"}]}
constraints = [
    Constraint("fetch_before_save", Before("fetch_data", "save_results")),
    Constraint("save_once", CalledNTimes("save_results", 1, "==")),
]
print(verify_trace(trace, constraints)["compliance_score"])   # 1.0

A trace is the list of tool calls an agent made. The score is the share of constraints the trace satisfies, weighted: 1.0 means all of them. Swap the two calls (save_results first) and fetch_before_save fails, so the score drops below 1.0. To refuse calls while the agent runs, instead of scoring afterwards, see enforcement.

More on the Python library

Your first rule

A rule file is YAML. This one makes the agent run the tests before it pushes, and run them again after any edit:

AGENTLTL.yaml
rules:
  - id: tests-before-push
    before: {first: pytest, then: git_push, since: [Edit, Write]}
    why: CI is slow; run the tests locally first.
    fix: Run pytest, then push.

Reading it:

  • pytest and git_push are tool names. A shell command is translated into one: git push origin main is git_push, and pytest is pytest (how that works). Edit and Write are the agent's own file tools.
  • before means git_push is allowed only once pytest has run.
  • since: [Edit, Write] means that run has to come after the last edit, so editing a file after the tests ran sends the agent back to run them again.

The agent sees why and fix when it is refused. In Claude Code, Copilot CLI, Mistral Vibe and Codex you can also write rules in plain words (/agentltl:rules never push to main, /agentltl-rules in Copilot, $agentltl:rules in Codex), and switch on tested ones from the library.

Next: what a rule can say.