Decawork

What is Decawork?

Decawork is a governance layer for the AI agents running inside a company. Teams route API requests and MCP tool calls through it, write policies in plain English, and a small model trained for their company approves or blocks each agent action in real time, based on the current task and past actions.

Top Features:

• Real-time approvals: a company-specific model checks each agent action before it runs.
• Agent registry: see every running agent with its owner, team, tools, and recent activity.
• Policy rules: write rules in plain English and choose when a person must approve.

Use Cases:

• Task-scoped access: let an onboarding agent read only the assigned new hire's profile.
• Agent inventory: track every internal agent built with Claude Code, Codex, or other stacks.
• Access revocation: remove agent permissions automatically once the assigned task is finished.

Who Can Use Decawork?

• IT teams: approve, monitor, and retire the agents employees build across departments.
• Security leaders: keep every agent action logged and tied to company policy.
• Engineering teams: ship internal agents faster because access and approvals are handled centrally.

Pricing

• Live walkthrough (free): book a demo to see approvals, policies, and the agent registry.
• Company plan (contact sales): custom pricing quoted for each company after a demo.

Pros and Cons

Pros:

• Context-aware decisions: the same action can be allowed for one task and denied for another.
• Any agent stack: connects agents built anywhere by routing API and MCP traffic.
• Hidden credentials: agents never hold raw keys, since credentials stay behind the gateway.

Cons:

• No public pricing: you must book a call before learning what the service costs.
• Early stage: a young two-person startup from the Summer 2026 Y Combinator batch.
• Setup effort: agents must route their traffic through the gateway before governance works.

FAQs:

1) What does Decawork actually do?

It trains a company-specific model that approves or denies agent actions.

2) Does it build AI agents?

No, it governs agents you already run, built with Claude Code, Codex, or others.

3) Can a person step in on risky actions?

Yes, policies can hold an action for human review until someone approves it.

4) Is it secure enough for enterprises?

It holds SOC 2 certification and shares security reports through a trust center.

5) How much does it cost?

Pricing is not published, so book a demo to get a quote.