by Phoebe Technology Limited
Open-source, self-hosted gateway where agents write TypeScript programs to call tools under Allow, Ask or Block rules; about 5,000 GitHub stars.
Pick Coral when agents must touch sensitive systems and you want per-operation approvals on infrastructure you control. It replaces hand-written API wrappers for each agent. The verdict: strong for engineering teams today, certified to ISO/IEC 27001:2022, though enterprise sign-on and audit logging are still coming and the benchmarks are the maker's own.
Agents save the programs that worked as custom tools, and each saved tool carries its own Allow, Ask or Block rule.
Saved custom tools and run traces accumulate inside a workspace, which makes leaving costly, but the gateway code is open source so the layer itself can be forked.
Main trade-off
The enterprise controls a regulated buyer asks for first are not shipped yet.
| Plan | Monthly price | What it includes |
|---|---|---|
| Open Source | Free | Coral tool execution layer, Agent permissions control plane, Shared agent workflows, Unlimited data sources, Unlimited agents, Unlimited queries |
| Enterprise | Custom | Everything in Open Source, SAML SSO and SCIM provisioning, Advanced RBAC and agent policy controls, Immutable audit log, Bring your own secrets vault, Custom provider integrations |
Open Source: free under the Apache 2.0 licence, self-hosted, with unlimited data sources, agents and queries; you bring your own LLM keys and your model provider bills you. Enterprise: custom pricing quoted on request, adding SAML SSO, SCIM, advanced RBAC, an immutable audit log and support SLAs; the README lists Enterprise as coming soon. Source: withcoral.com/pricing.
Coral is an open-source, self-hosted AI gateway in the agent gateways category. Agents write short TypeScript programs that run in a V8 isolate, and Coral checks every call against Allow, Ask or Block policy before it adds credentials. The code is Apache 2.0 licensed and the latest release is 1.8.0.
Agent-written programs
Agents call tools through short TypeScript programs that run sandboxed, so a program can only reach what Coral lets it call.
Allow, Ask or Block rules
You set a rule for each provider tool and each custom tool, and an approval request shows the exact call and its arguments.
Saved custom tools
A program that worked becomes a private tool the agent finds in the same search as provider tools, and an owner can share it workspace-wide.
Providers from OpenAPI or MCP
Add a service from the catalog, an OpenAPI document or an MCP server, and agents find its tools in one search.
Per-user accounts
A provider can run as the caller's own account, so teammates share tools without sharing API tokens.
Run traces
Every run is attributed to the person or agent that started it and records the goal, the tools used and the outcome.
Desktop app or shared server
The desktop app runs on macOS (Apple silicon), Windows (x64) and Linux (x86_64); the shared server needs an OIDC identity provider and SQLite or Postgres.
closeStaleTicket changes one field on tickets idle for 30 days, so security approves one action instead of a whole API.| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| -- | Yes | Web · Mac · Windows | 3.6 | |
| $15/mo | Yes | Web | 3.2 | |
| $29/mo | Yes | Web | 3.8 | |
| $120/mo | Yes | Web | 3.5 |
Apache License 2.0 for the open-source edition; the ee/ directory holds Enterprise Edition code under its own license
The open-source edition is free under the Apache 2.0 licence, with unlimited data sources, agents and queries, and you run it yourself with your own LLM keys. Enterprise is priced by quote and adds SAML SSO, SCIM, advanced RBAC and an immutable audit log, though the README lists it as coming soon. No per-seat or per-call price is published.
Yes. The open-source edition has no limits on users, agents or volume, and the desktop app lets you try it on a laptop before any server rollout. Your costs are the model keys your agent already uses and, for a shared server, the host, identity provider and database you run yourself.
Any agent or app that supports MCP. The README shows claude mcp add for Claude Code and codex mcp add for Codex, and points to a guide for Cursor, VS Code and OpenCode. You copy your workspace's MCP URL from the Coral UI and paste it into the client.
A typical MCP gateway decides which tools an agent can see, so a tool such as update_record can change any record it allows. Coral adds rules per operation, human approval and custom tools with fixed arguments. The vendor's own tests compare it with Kong, Composio and Microsoft's gateway.
Coral calls your SaaS APIs and databases directly, so nothing is copied into a warehouse first. It keeps a small local store of metadata, custom tools and run records that you own, and the vendor says nothing is sent to it when you run Coral yourself. Only the results an agent asks for reach its model.
Coral is an open-source, self-hosted gateway from Phoebe Technology Limited (the repository is withcoral/coral) that sits between AI agents and the systems they work on. The team says it began while building an AI site reliability engineering agent that needed broad permissions and many tokens, and started from scratch on every incident. The repository was created in April 2026, reached version 1.8.0 on 13 November 2026 and has about 5,000 GitHub stars. A Show HN post on 16 November 2026 collected 51 points.
Agents call tools by writing short TypeScript programs and sending them through one exec tool. Each program runs in a fresh V8 isolate with no network access, no credentials and no imports. Coral checks every tool call against workspace policy, adds the credentials, then forwards the call to the provider, and it can pause the run until a person approves. When a program works, the agent saves it as a private custom tool, and a workspace owner can publish it as a shared tool for everyone.
It fits teams that let agents act on real systems: site reliability engineers who allow a rollback of named services only, support teams whose agent issues refunds with defined escalation, security teams isolating devices within approved groups, and analysts who want a custom tool to aggregate warehouse, CRM and billing data before the agent sees it. Agents connect over MCP, including Codex, Cursor, GitHub Copilot, ChatGPT and Gemini. You can start on the desktop app for macOS, Windows or Linux, or run a shared server with Docker Compose or Kubernetes.
Some buyers should pick something else. Composio hosts a large catalogue of ready-made app connections, and Executor hosts the apps and schedules an agent writes, so neither asks you to run a server. Clawvisor approves agent work by task and vaults credentials. Kong and Microsoft also publish MCP gateways that the vendor benchmarks against, and neither is linked here. Teams that need SAML single sign-on, SCIM provisioning or an immutable audit log today cannot get them: Coral Enterprise is listed as coming soon. For the wider shelf, see AI gateways and best AI agent builders.
The 1.8.0 release let operators define secret stores for credentials, and the 1.6.0 release of 10 November let agents save and delete saved functions from inside a program. Coral 1.0 replaced its earlier SQL interface with TypeScript programs and starts from fresh state rather than converting what version 0.x stored.
Our call: choose Coral if your team can run a server and wants per-operation approvals on agent actions. Choose Composio or Executor if you would rather not run one, and wait if you need single sign-on or an audit log this quarter.
Product facts on this page (pricing, plans, integrations, compliance) are read from Phoebe Technology Limited's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.