Pinned Tweet
Coding agents on auto are great right up until one tries a production deploy or an npm publish.
Marrow sits in your agent's pre-tool hook. Routine work flows. On the evaluation and Business plans, risky actions wait for your yes. Every hold and every answer is recorded.
We're opening it up to early users this week. 🧵
1
3
118
AI agents are getting smarter.
The problem? They’re still free to act without a real control layer.
Most teams have:
→ Better models (the brain)
→ Better harnesses (the hands)
→ Logs after the fact
What they’re missing is the gate in between.
Marrow checks risky actions *before* they run, holds the ones that need approval, and leaves proof the next agent can actually reuse.
Result:
• Fewer surprises
• Accountable outcomes
• Agents that get better over time instead of drifting
We’re not a harness. We’re the control layer that sits in front so the work stays safe and the fleet improves.
30-day full evaluation, no card.
🤖 Made with AI
1
22
We’re not a harness. We’re the layer that sits in front so you stay in control and the work gets better.
11
Anthropic is changing the game.
OpenAI, and others need to catch up to model performance, and cost efficiencies Claude provides before Fable 5.5 drops and extends their rank 1 frontier status even further.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
24
OpenAI and Anthropic told California how to stop risky AI.
OpenAI: shutdown controls, like isolating an agent or cutting tool access. Anthropic: circuit breakers, down to one agent.
Greyhound Research: "Every assigned duty needs an executable control."
csoonline.com/article/424212…
1
1
32
With an account, Marrow checks every risky action before it runs and keeps proof.
getmarrow.ai
9
Agent tried a command and it didn't run? In Claude Code, ctrl+o shows Marrow's reason.
The free seatbelt refuses seven kinds of protected action on your machine. No account, no key, no network call.
/plugin marketplace add getmarrow/plugins
/plugin install marrow@getmarrow
1
41
Marrow AI is on Product Hunt today: producthunt.com/products/mar…
Rules, approvals and proof for your AI agents. The free seatbelt needs no account and no key, and stops 7 kinds of risky agent action on your machine.
Tell us the first time it gets in your way.
17
18
174
Correction: we launch on Product Hunt on Thu 19 Nov (12:01 AM PST), not today. Marrow is already live and free to try.
It checks what your AI agents are about to do before they run. getmarrow.ai
12
Marrow AI is one control layer for AI agents on every model: Claude Opus 5.5, GPT-6 Astra, Grok 4.7, Gemini 4.
Free seatbelt for Claude Code, no account or key: stops force-pushes to protected branches and prod deploys. Not native Windows yet.
github.com/getmarrow/plugins
15
15
204
Marrow AI retweeted
Marrow is up on Show HN.
It's a seatbelt for AI coding agents. It runs locally, needs no account, and stops moves like a force push to main before they land.
Works on macOS and Linux today, Windows through WSL. Honest feedback in the thread is very welcome.
news.ycombinator.com/item?id…
Tomorrow we're on Hacker News with Marrow AI: the control layer for AI agents.
Today: try the free seatbelt and tell us the first moment it got in your way. Every unneeded prompt is a bug for
us.
getmarrow.ai
2
87
What does Marrow AI actually do on each host? The Works-with table on our site now says, limits included.
Ten hosts: where it stops an action, where it only advises, and what we have not yet verified in the live app.
getmarrow.ai/#works-with
2
22






