The brain for AI companies. In your terminal.
Checks what an agent's about to do against your rules before it does it. Everything else lives as files in your own repo, and approval is a merged PR.
curl -fsSL gntai.dev/install.sh | sh$ npm install -g @gnt-ai/cli$ gnt login✓ Logged in. Credentials saved to ~/.gnt/credentials.json$ gnt prebrain# scans this repo, asks a few questions, drafts your first rules✓ Opened PR: https://github.com/acme/rules/pull/412 (6 rules)# merge it on GitHub. that merge is the approval.
What is gnt.ai?
gnt.ai is a git-native rules layer for AI agents. Before an agent acts, it checks the action against your team's approved rules. It can also search or fetch those rules directly. Every rule is a file in your own repo, reviewed and merged like code.
Enforces, not just retrieves check_action is the one enforcement call: it runs right before the agent's next step and returns allowed, blocked, or needs_human, citing the rule that decided it.
MCP-native One endpoint, five tools: check_action, search_rules, get_rule, list_skill_packs, get_skill_pack.
Git-native Rules live as files in your own repo. Versioned, diffed, and reviewed like code.
Human-approved A merged pull request is the approval. Same review flow your team already uses for code.
Drafts your first rules gnt prebrain scans your repo, a docs directory, and a Notion export, asks a few questions about your company, and opens real pull requests. Editable starter packs cover you if there's not much written down yet.
Pulls from where policy already lives Slack's /brain command, a webhook (Zapier and friends), or Notion and monday.com over MCP all draft rules the same way. Nothing skips review, they just skip the copy-paste.
Finds the gaps gnt gaps surfaces every question an agent asked with no covering rule, so you know exactly what to write next.
Flags what's stale gnt stale tracks rule freshness and opens a pull request to refresh or deprecate what's due, before it quietly goes wrong.
Catches contradictions A nightly sweep finds rules that conflict and opens a pull request proposing which one should defer to the other, plus a GitHub issue.
Any agent Claude, GPT, or your own MCP-compatible client. The endpoint doesn't care.
How it works
- 1.
Draft. gnt prebrain gets your first rules from your repo, docs, and Notion in one run. After that, new rules can come from Slack, a webhook (Zapier and friends), Notion or monday.com over MCP, or written by hand.
- 2.
Review. gnt review turns anything drafted outside prebrain into a real pull request against your own rules repo. prebrain opens its own PRs directly.
- 3.
Merge. A human merges it on GitHub. That merge is the approval. No dashboard, no separate sign-off screen.
Built to stay out of your data
Rules live as files in your own git repo, not a database we control. Nothing is approved, or served to an agent, until a human merges the PR.
gnt prebrainand its walkers (repo scan, docs directory, Notion export, and the Notion / monday.com MCP-in connectors) are CLI-driven: raw sources are processed on your device, anything sent to a cloud model passes a local privacy gate, your sources never touch gnt's servers, and we never train on your data.
A webhook or the Slack /brain commandworks differently: that content (a monday.com comment, a Slack message) has no device of yours to run on, so it reaches gnt's server directly. It's masked there, on arrival, before anything is stored, by a server-side privacy gate, and that masking is permanent: there's no way to recover the original text afterward. See how it works.
Pricing
One tier. No metering, no per-seat pricing. Every org gets a 14-day trial starting the moment they sign up, no card required until you run gnt billing to subscribe.
Frequently asked questions
Approval is a merged PR. Nothing simpler.
No dashboard to review in, no separate sign-off screen. Just the git workflow your team already trusts.