HEXIS CLI
HEXIS compiles a SKILL.md into a state machine: code fixes the order of steps and the model reasons only inside each state.
- 136 ★ on the repository
- Last commit
- GPL-3.0
Extended access
Why this level
- The `--executor local` mode runs `bash` commands on the user's machine.
- Skill text and traces go to the configured model endpoint together with the key from the environment.
Install
- Works in
Manual install
pip install hexis-agentFirst install command from the README. The package imports as `hexis`, the command is `hexis-agent`, Python 3.11 or newer is required.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The `--executor local` mode runs `bash` commands on the user's machine.
- Skill text and traces go to the configured model endpoint together with the key from the environment.
This is third-party code. Review the repository files before installing.
What it does
HEXIS is a Python command-line tool that turns a `SKILL.md` document into an extended finite state machine. The machine stores a current state and variables, runs the state's action and picks the next step from transition guards. The model reasons and generates inside states while the program keeps the order of operations. The `compile` command builds the machine through any OpenAI-compatible model and runs static checks. `update` folds execution traces into the machine and accepts a change only if every previously accepted trace still replays. `run` executes the machine on a task. Each build also gets `GUIDE.md` and `PROMPT.md`, a prompt that lets any tool-using agent walk the machine step by step.
Who it is for. Skill authors whose agents skip or reorder mandatory steps, and researchers who need reproducible execution of instructions.
Good fit when
- An agent runs a skill with a strict order of steps and keeps skipping checks.
- Execution traces have piled up and the procedure should be refined from them without hand-editing text.
- A weaker model gets the procedure and code must guarantee the order of operations.
- The team wants a readable map of a skill: states, transitions and variables in JSON.
Not a fit when
- The skill is short and loose, and a rigid order of steps would hurt it.
- No OpenAI-compatible model is available to build the machine.
- You want a ready Claude Code plugin without Python or a separate runner.
Example request
Compile my skill into a machine, show the state diagram and run it on a test task.Limitations
The project is in alpha under GPL-3.0. Building and running need an OpenAI-compatible model and a key in an environment variable, and native tools need OpenCode. The `--executor local` mode runs `bash` on your machine, so run machines only from trusted sources and in a separate working folder. Machine quality depends on the model that builds it.
How to disable. Remove the package with `pip uninstall hexis-agent` and delete the build folders. Remove the key from the environment variable or `.env` separately.
FAQ
Which command builds a machine?
`hexis-agent compile --skill path/to/skill --out build/` with the `--model`, `--base-url` and `--api-key-env` flags. The model drafts the machine and redrafts it until static checks pass.
Can an update be checked without calling the model?
Yes. `hexis-agent update --build build/ --traces traces/ --show 3` previews the result, makes no model calls and writes nothing.
How does another agent use a finished machine?
Through `PROMPT.md` in the build. It is a system prompt that lets an agent with tools execute the machine step by step.
What happens if the machine is only partly learned?
The fallback state retries the most recent tool step and then hands the task to ordinary execution of the skill.
Does the README example need internet?
No. The `table_clean` example ships with scripted tools and a scripted model and runs offline.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI that compiles a SKILL.md file into a standalone Python agent with tools, knowledge-base retrieval and its own runner.
A meta-skill that folds rules scattered across AGENTS.md, CLAUDE.md and Cursor files into one project skill with task routes.
A meta-skill that splits a task into stages, picks a profile and runs multi-agent work in Codex and Claude Code.
A CLI that runs your agent on tasks and revises its skills, keeping a change only when held-out tasks confirm the gain
A self-hosted git store for skills, tools and context: agents fetch them over MCP and every change goes through review
A skill with 25 commands to audit and improve the prompts, tools and agent designs of a project in Cursor, Claude Code and Codex
README in short
The package accompanies an arXiv paper. The repository ships four compiled machines for data analysis, mathematics, corpus questions and spreadsheet editing, plus tests that need no network.