
Coral
Execution rails for agents
About Coral
Coral is an open-source execution layer for AI agents, built by Phoebe Technology Limited and licensed Apache-2.0. Instead of handing an agent a broad scope and a long list of MCP tools, you turn any OpenAPI document or MCP server into a provider and set policy per operation: allow it, require a human approval, or disable it. Agents then write short TypeScript programs that Coral runs in a sandbox, adding your credentials and checking every call against your rules. Credentials, data and run history stay on your machine or your own servers.
You install the desktop app or self-host Coral for a team, then connect providers, which are any OpenAPI document or MCP server, into workspaces that group employees and agents with their own rules. When an agent needs to act it searches for the operations it needs rather than loading them all, writes a short TypeScript program, and Coral runs it in a sandbox with your credentials attached. Actions you marked as needing approval surface in the agent's own interface. Programs that work are saved as reusable custom tools, and every run is recorded.
- •Per-Operation Policy - Each operation can be allowed, set to require human approval, or disabled, so permissions describe actions rather than broad scopes.
- •Providers From OpenAPI or MCP - Any OpenAPI document or MCP server becomes a provider your agents reach through one server, with your credentials attached.
- •Reusable Custom Tools - Programs that succeed are saved to a shared library, so agents reuse proven code instead of writing it again each task.
- •In-Agent Approvals - People approve the actions that need approval inside the agent's own interface, such as Claude or Codex, and Coral logs each one.
- •Local-First With Run History - Credentials, data and run history stay on your machine or your servers, and every program and the calls it made are recorded.
Teams connecting agents to systems where a mistake costs something, such as billing, customer records or production infrastructure, and who need to say what an agent may do rather than only what it may reach. Security reviewers get an audit trail and local credential storage. It is more machinery than you need if your agent only reads public data.
Pricing
Custom pricing, contact sales
- Open Source$0/mo
- EnterpriseContact sales

























































