What Is Coral (2026)? Open-Source Gateway for AI Agents
By Antigravity Skills editorial team · Last updated
Short answer: Coral is an open-source, self-hosted gateway between AI agents and systems like GitHub or Slack. An agent writes a short TypeScript program; Coral runs it in an isolated sandbox, checks each call against your policy, and can save a working program as a reusable custom tool. The core is Apache 2.0.
Unofficial guide. Antigravity Skills is not affiliated with Coral or its makers. Every fact below comes from the withcoral/coral repository, the quickstart page and the Show HN thread, read on November 18, 2026. Numbers such as speed and cost savings are the vendor's own claims, not our measurements. Confirm details on the official pages before relying on them.
What is Coral?
"Coral" is a crowded name: Google has had a Coral hardware platform, and several unrelated companies use it too. A commenter on Hacker News raised exactly this confusion, and one of the founders agreed a suffix could help. This page is only about withcoral/coral, the agent gateway on GitHub.
The README's one-line pitch is "execution rails for agents". In its own words, Coral is a gateway you host yourself that agents reach over MCP (Model Context Protocol). A few key terms from the docs:
- Provider: one service your agents can use, such as GitHub or Linear. You pick one from Coral's catalog, or bring your own OpenAPI document or MCP server.
- Custom tool: a program an agent wrote, which worked, and which was saved so it can be called again in one step.
- Policy: for each provider tool and custom tool you set Allow, Ask or Block.
- Workspace: the boundary that decides which providers and tools agents can reach. Only workspace owners can add providers or change policies.
The project started inside an AI SRE (incident investigation) product, according to the founders. They say the agent needed broad permissions and many tokens, and began from scratch each time. Coral was built so the agent "remembers the route".
How Coral works
The README describes a short loop.
- The agent connects to Coral's MCP server and uses an
exectool to send a TypeScript program. - The program calls
coral.searchto find available tools and read their TypeScript signatures, then a second program does the actual work. - Each program runs in a fresh V8 isolate with no network access, no credentials and no imports.
- For every tool call, Coral checks your workspace policy, attaches the credentials itself, and forwards the call to the provider. A policy can pause the run until a person approves.
- If the program succeeds, the agent can save it as a private tool, called as
coral.private.<name>(...). A workspace owner can publish it as a shared tool undercoral.shared.
Two details from the README explain why this design matters. A program can filter large API responses and return only what the task needs, so less raw data enters the model's context. And because the run happens on the Coral server, it continues even if the agent disconnects.
The README's own example finds open GitHub pull requests with no reviewer and posts them to Slack. Posting needs approval, so the run waits for you, and only the pull request numbers go back to the agent.
Why custom tools are the selling point
The README argues that most MCP gateways only decide which tools an agent can see, which is all or nothing. A custom tool is code, so it can hold a rule that Allow and Block cannot, for example a closeStaleTicket tool that changes one field and only on tickets idle for 30 days. You then set Allow, Ask or Block on that tool like any other. Coral still checks every provider call inside it.
What the README claims, and what we could check
This is our own breakdown of the main claims. "Source says" means the claim is stated by the project; we did not run Coral or reproduce any benchmark.
| Claim | Where it appears | Our status |
|---|---|---|
| Open-source, self-hosted gateway | README, HN post | Repository is public; GitHub reports an Apache-2.0 license and Rust as the main language |
| Apache 2.0, no limits on users, agents or volume | README, HN post | Stated by the project; an ee/ folder holds Enterprise code under a separate license |
| Runs as a desktop app or a shared server | README | Stated; desktop builds listed for macOS (Apple silicon), Windows (x64) and Linux (x86_64) |
| Programs run in a V8 isolate with no network or credentials | README, HN post | Stated by the project; we did not audit the code |
| 29% faster, 31% cheaper on GitHub, Linear and Slack tasks | HN post | Vendor figure, with Opus 5.5; the README adds 37% fewer tokens across 34 tasks. Not independently reproduced |
| ISO/IEC 27001:2022 certified | README | Stated by the project; certificate not checked |
| Enterprise features (SAML SSO, SCIM, audit log) | README, HN post | Described as "coming soon"; not part of the open-source edition |
How to get started
The README lists four steps. We have not run them, so treat this as a summary of the official instructions rather than a tested walkthrough.
- Install Coral. To use it alone, download the desktop app from the official download page. For a team, run a shared server with Docker Compose or Kubernetes; the README says you need an OIDC identity provider, a reverse proxy for TLS, and SQLite or Postgres.
- Add a provider. The quickstart says only workspace owners can do this. By default, tools that read data run immediately and everything else needs approval. The quickstart does not define which tools count as "reading", so check each provider's tools.
- Connect your agent. Copy the MCP URL from the Coral UI and add it to your MCP client. The README shows
claude mcp addfor Claude Code andcodex mcp addfor Codex, and links to instructions for Cursor, VS Code and OpenCode. - Ask something that uses the provider, for example "What are my open pull requests?"
Coral is not an agent itself. It does not replace Claude Code, Codex or Antigravity; it is the layer your agent talks to. If you are new to those agents, start with our guide to installing skills in Antigravity, Claude Code and Codex.
Practical advice from the founder
In the Hacker News thread, co-founder James answered a question about deployment. Paraphrasing his three points:
- Start with the desktop app before investing in a shared server, which needs your own identity provider.
- Do not design custom tools up front. Let agents discover what they need, and keep a human reviewing additions and updates.
- Decide whether users authenticate to each provider with a shared token or as themselves. Most prefer the latter, which means each person completes a one-off OAuth flow per provider.
Caveats and open questions
- Approvals are awkward across agent apps. The founder said Coral first used MCP elicitation, but agent apps handle it inconsistently: one is said to answer "decline" automatically, while under some settings another does the opposite. Coral now returns an approval URL in the tool response and, where supported, shows an approval dialog as an MCP App. Expect to open the Coral UI to approve calls.
- Young project, fast changes. The repository was created in April 2026, the latest release we saw was v1.8.0 on November 13, 2026, and v1 replaced the earlier SQL interface with TypeScript programs without converting old 0.x state. The README has an upgrade section for that.
- Benchmarks are the vendor's. The savings figures come from tasks the project chose, on three services. Your results will depend on your tasks and models.
- Launch-thread noise. A few new accounts commented on the launch thread, and one commenter flagged it. The founder replied that early preview users had been told about the launch and some made new accounts. Weigh the praise in that thread accordingly; the concrete technical answers are the useful part.
- Reading is not auditing. Coral holds credentials for your systems. Review its source, permissions and update process like any other infrastructure before connecting production accounts.
Where Coral fits with other agent tools
Coral sits below your agent, not beside it. Compare it with the agent runtimes and skill formats we cover elsewhere: OpenClaw is a self-hosted assistant with its own gateway, and gstack is a skill pack that changes how an agent works. Skills tell an agent how to do a task; Coral controls which systems it may touch and how. The sources do not benchmark Coral against other MCP gateways beyond the vendor's own claim, so we do not rank alternatives here. To browse ready-made skills, see the Antigravity Skills library.
FAQ
What does Coral do?
It is a self-hosted gateway between AI agents and your tools. Agents write TypeScript programs that run in a sandbox, Coral checks each call against your policy, and working programs can be saved as reusable custom tools.
Is Coral free and open source?
The core is Apache 2.0 with no limits on users, agents or volume, according to the README. Enterprise features such as SAML, SCIM and an immutable audit log are planned under a separate license.
Is Coral the same as Google's Coral?
No. This is withcoral/coral, an agent gateway. Google's Coral was a separate hardware platform with the same name.
Which agents work with Coral?
Any app or agent that supports MCP, per the README. It shows commands for Claude Code and Codex and links to setup for Cursor, VS Code and OpenCode.
Related on Antigravity Skills
- How to install skills in Antigravity, Claude Code and Codex
- OpenClaw: the self-hosted AI agent
- gstack
- All skills