ai-engineering CLI
ai-engineering installs hooks that deny risky agent calls and write a receipt for every verdict into your repository
- 65 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- Hooks are installed into hosts and git and deny agent calls, and a faulty rule can stall a workflow
- `init` writes contract files and git hooks into the repository, so review the changes before committing
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/ai-engineeringDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Hooks are installed into hosts and git and deny agent calls, and a faulty rule can stall a workflow
- `init` writes contract files and git hooks into the repository, so review the changes before committing
This is third-party code. Review the repository files before installing.
What it does
ai-engineering is a package with an `ai-eng` command that turns rules a repository already holds into code that runs on every agent tool call. Host hooks send the call to the CLI, which answers in that host's dialect: allow, deny or a rewritten command. `ai-eng` itself never calls a model and sends no prompts or source to a service. `ai-eng init` asks which agents to govern, installs the skill canon on your machine and writes `AGENTS.md`, `DECISIONS.md`, `.ai-engineering/config.toml` and git hooks into the repository. Then `ai-eng doctor` runs its checks and fires one real adversarial payload down the chain to confirm a deny came back. The guards refuse `git commit --no-verify`, a silenced linter, an edit to the approved contract and loops of repeated calls, and they stop a file that argues with the model before it is read. A guard that crashes or cannot decide denies. The only bypass is written into `.ai-engineering/overrides.toml` with a reason and an end date, and `doctor` reports the time left. Every verdict gets a receipt on disk and in git. Hosts fall into tiers: core with deny and rewrite, experimental, best-effort and skills-only with no interception.
Who it is for. Developers and teams that let Claude Code, Codex and other agents run commands and want repository rules enforced on every call.
Good fit when
- An agent tries `git commit --no-verify` and you need a denial with a clear explanation
- `AGENTS.md` rules get forgotten in long sessions and need to become a check before the call
- Every guard decision has to stay as a receipt in the repository for later review
- An exception to the rules has to be written down with a reason and an expiry and seen in `doctor`
Not a fit when
- The machine has no Bun 1.4.4 or newer: `ai-eng` runs TypeScript source through Bun
- You need a network firewall or gateway for agent traffic: the tool works on host calls
- Your host is in the skills-only tier: there is no interception on it
Example request
Run ai-eng doctor in this repository and show which hosts passed and which guards are activeLimitations
- Bun 1.4.4 or newer is required because `ai-eng` runs its TypeScript source through it; without Bun the launcher stops with a clear message.
- Hooks do not run in an untrusted folder, so you must trust the workspace in your host yourself, the one step `ai-eng` cannot do.
- Coverage depends on the host: the core, experimental, best-effort and skills-only tiers behave differently, and `doctor` reports it per host.
- At most once a day `init` and `doctor` may read package registry metadata; set `AI_ENG_NO_UPDATE_NOTICES=1` to turn it off.
How to disable. `ai-eng uninstall` reverts what it installed and keeps your files, and `ai-eng config --remove <host>` turns one host off. Remove the package with `bun remove -g ai-engineering` or `npm uninstall -g ai-engineering`.
FAQ
How do I install ai-engineering?
Run `bun add -g ai-engineering@latest && ai-eng init` or `npm install -g ai-engineering@latest`. Bun 1.4.4 or newer is required.
What does ai-eng init write into the repository?
`AGENTS.md`, `DECISIONS.md`, `.ai-engineering/config.toml` and git hooks. Changes are left for your review, and the command can be run again.
How do I check the guards work?
Run `ai-eng doctor`: it runs every check it knows and fires one real adversarial payload down the chain, then reports whether the deny came back.
How do I bypass a wrong denial?
Add an entry to `.ai-engineering/overrides.toml` with `reason` and `until`. An expired entry re-arms the guard.
Does it call models or a cloud?
No. There is no hosted service or daemon and `ai-eng` itself never calls a model. The single exception, reading package registry metadata at most once a day, can be switched off with an environment variable.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI with policies that intercepts dangerous AI agent commands like rm -rf, force push or reading .env before they run
HackAgent tests your AI model or agent for jailbreaks, prompt injection and tool misuse and reports a measured attack success rate
An execution firewall for agent actions: decides ALLOW, DENY or ESCALATE and writes a signed receipt that can be verified later.
Agents Shipgate shows in a pull request how an agent permissions and tools changed before the changes reach main
Hook rules with no model in the loop stop dangerous commands and secrets in code, and every block goes to a local log.
AgentGuard checks a coding agent's actions before they run and scans third-party skills for secrets, backdoors and injections.
README in short
The README states the goal (a contract, denials, receipts), shows the numbers (five guards, one output filter, zero services), explains the idea and security boundary, installation, quick start, verdicts and the list of blocks, then skills, surfaces and commands.