What is Marrow AI?
Marrow AI is a governance layer for AI agents, delivered through plugins for Claude Code, Cursor and Codex, along with hooks and an MCP server. It checks risky actions that pass through its controls and, with an account, can hold them for approval and keep proof. Teams can review prior outcomes and measure reliability across a tenant-scoped agent fleet. Claude Code and Cursor users also get a local keyless seatbelt that works without an account.
Marrow AI Features:
Agent Plugins: Claude Code and Cursor plugins add Marrow hooks, an MCP server and a skill, while Codex offers hook-enabled and advisory-only plugins.
Pre-Action Checks: Marrow checks policy, risk and relevant prior outcomes before an agent takes an action.
Keyless Seatbelt: Claude Code and Cursor hooks can locally stop seven types of risky actions without an account or key, including production deploys and credential reads.
Approval Holds: With an account, Marrow can hold risky actions for approval when they pass through its hook, runner or MCP gate.
Outcome Proof: Teams can close meaningful work with outcome evidence and retain reusable proof.
Fleet Measurement: Marrow measures reliability, avoided failures and improvement across a tenant-scoped agent fleet.
Marrow AI Benefits:
Human Checkpoints: Teams can put review between an agent and a risky operation instead of relying on the agent to decide whether to proceed.
Local Safeguards: Teams using Claude Code or Cursor can add basic protections before connecting an account or configuring a key.
Reusable Evidence: Recorded outcomes give teams material to refer back to when assessing similar work or reviewing an agent’s decisions.
Fleet Visibility: Reliability and improvement measures help teams see how agent performance changes across their fleet.
Low-Commitment Evaluation: The 30-day evaluation requires no card and includes up to 10 agents and 3,000 decisions per month.
Use Cases:
Coding Agent Setup: Add governance hooks to a Claude Code or Cursor workflow so risky actions can be checked before the agent carries them out.
Production Deploys: Require review when an agent attempts a production deployment or another protected operation.
Credential Protection: Use the local seatbelt in Claude Code or Cursor to stop an agent from reading credentials or project secret files.
Agent Fleet Review: Measure reliability and avoided failures when managing multiple agents within a tenant.
Governance Evaluation: Try the controls with a small agent setup before deciding whether to use them in production.

