Coral 开源智能体网关发布:智能体可自写 TypeScript 工具并复用
Show HN: Coral – Open-source gateway where agents write their own tools
Coral 是一个开源、自托管的智能体网关,位于智能体与其操作的系统之间,智能体通过编写短 TypeScript 程序调用工具,成功的程序可保存为自定义工具复用。
原文给出了自定义工具复用与逐调用策略校验的具体机制,以及 34 项任务上的 token、成本与耗时对比数据。
Execution rails for agents.
Coral is an open-source, self-hosted gateway that sits between your agents and the systems they work on. Agents call tools by writing short TypeScript programs, and save the ones that work as custom tools to reuse. Coral checks every call against your policy before it adds credentials and forwards the call.
Quickstart · Download · Demo video
In this example, an agent posts unreviewed pull requests to Slack for the first time. The agent searches for the GitHub and Slack tools and writes code that calls them. It cannot post to Slack without approval, so the user approves the post. The agent then saves its code as a custom tool. The next day, the agent runs the same task by calling that custom tool.
Control what agents can do. Most MCP gateways decide only which tools an
agent can see. That is all or nothing: if an agent can call update_record, it
can update any record in any way the tool allows. A custom tool is code, so it
can hold a rule that Allow and Block cannot. For example, closeStaleTicket
changes one field, and only on tickets with no activity for 30 days. You set
Allow, Ask, or Block on the custom tool like any other tool. Coral still checks
every provider call inside it against your policy.
Less duplicated agent work. Agents start every task from scratch. They re-read tool schemas, retry failed calls, and spend tokens rewriting code for steps they got right yesterday. With Coral, agents save the code that worked as custom tools, and call them again in one step. Workspace owners can share those tools with the whole team.
Thousands of developers at more than 450 companies use Coral for safer, more efficient agent work. They use it in customer support, SecOps, business intelligence, SRE, and more.
Get started
-
Install Coral.
- To use Coral yourself, download the desktop app from the download page. It runs on macOS (Apple silicon), Windows (x64), and Linux (x86_64).
- To use Coral with your team, run a shared server with Docker Compose or Kubernetes. You need an OIDC identity provider, a reverse proxy for TLS, and SQLite or Postgres.
-
Add a provider. A provider is one service your agents can use, such as GitHub or Linear. Pick one from the catalog, or bring your own OpenAPI document or MCP server. By default, tools that read data run right away and everything else needs your approval. See Add a provider.
-
Connect your agent. Copy your workspace's MCP URL from the Coral UI and add it to your MCP client:
claude mcp add --scope user --transport http coral <your-mcp-url> # Claude Code codex mcp add coral --url <your-mcp-url> # Codex
For Cursor, VS Code, OpenCode, and other clients, see Connect an MCP client.
-
Ask a question. Ask your agent something that uses that provider, such as "What are my open pull requests?"
Why Coral
We started Coral while building an AI SRE. To investigate an incident, the agent had to connect the dots across an ocean of telemetry and undocumented relationships between internal and external systems. It needed broad permissions and a lot of tokens, and it started from scratch the next time.
We built Coral so that the agent remembers the route. Once a task succeeds, the agent saves the steps that worked as a custom tool. Next time, it finds that tool in the same search as provider tools such as Datadog and Slack. Coral started with SRE. Teams now use it for agent work across the organization:
- SRE. The agent investigates alerts. Custom tools let it roll back only the services you name.
- Customer support. The agent issues refunds, and escalates to a person in the cases you define.
- Business intelligence. Custom tools aggregate warehouse, CRM, and billing data, and pass only the metrics to the agent.
- SecOps. The agent uses custom tools to isolate devices within approved groups.
What you get:
- Precise control. Set Allow, Ask, or Block for each provider tool and each custom tool. An approval shows the exact call and its arguments, so Security approves one refund lookup, not the whole payments API. See Roles & Permissions.
- Shared tools. You can set each provider to use the caller's own account, so teammates share tools without sharing API tokens. Read how Slackbot uses this.
- Cheaper models, same accuracy. With custom tools, a smaller model performs like a bigger one. For many tasks, this means Opus 5.5 accuracy at Haiku 5.5 cost.
- Faster, leaner agents. Across 34 tasks on GitHub, Linear, and Slack with Opus 5.5, custom tools cut tokens by 37%, cost by 31%, and time by 29%. See every task.
- Ahead of other MCP gateways. In blind benchmarks of 100 enterprise tasks, agents were more accurate and cheaper with Coral than with other MCP gateways. See the benchmarks.
- Training data you own. Each run leaves a structured trace of the goal, the tools used, and the outcome. The trace works well for post-training your own model.
How Coral fits your stack:
- Runs on your infrastructure. Coral is written in Rust. It runs as a desktop app on your laptop or as a shared server in your own cloud, with zero data egress.
- Open source. Coral is licensed under the Apache License 2.0, with no limits on users, agents, or volume. Everything in this README outside Coral Enterprise is in the open-source edition.
- Any agent, any system. Coral works with any AI app or agent that supports MCP. You can connect anything with an MCP server or OpenAPI document, including more than 5,000 databases and SaaS tools.
How Coral works
Agents write programs. Coral builds on code mode, and adds custom tools, policy controls, and managed identities. To use tools, an agent writes a short TypeScript program. One program can call several tools, filter the results, and return only what the task needs. Less raw data passes through the model's context.
Coral runs them on rails. Each program runs in a fresh V8 isolate with no network access, no credentials, and no imports. Coral checks every tool call against your workspace policy, then adds the credentials and forwards the call to the provider. A policy can also require a person to approve the call. Coral attributes every run to the person or agent that started it.
graph LR
Agent["Your agent"] -->|"exec: run a program"| MCP["Coral MCP server"]
You["You"] -->|"providers, policy, approvals"| UI["Coral UI"]
subgraph Coral["Coral (your laptop or server)"]
MCP --> Run["Program runs in a V8 isolate"]
Run <--> FN["Custom tools (private and shared)"]
UI --> Gate["Policy check and credentials"]
Run --> Gate
Gate -.->|"approval request"| UI
end
Gate --> GH["GitHub"]
Gate --> SL["Slack"]
Gate --> Own["Your OpenAPI or MCP service"]
Loading
The agent sends Coral programs through the exec tool. A program calls
coral.search to find the tools the agent can use and read their
TypeScript signatures. Then the agent sends a program that does the work.
This one finds open pull requests that have no reviewer and posts them to
Slack:
const { data: pulls } = await coral.providers.github.pulls.list({ owner: "withcoral", repo: "coral", state: "open", }); const unreviewed = pulls.filter((pull) => pull.requested_reviewers.length === 0); await coral.providers.slack.chat.postMessage({ channel: "#engineering", text: unreviewed.map((pull) => `#${pull.number} ${pull.title}`).join("\n"), }); return unreviewed.map((pull) => pull.number);
If posting to Slack needs approval, the run pauses there until you approve it. Only the pull request numbers go back to the agent, not the full GitHub response. The run happens on the Coral server, so it continues even if the agent disconnects.
What works becomes a custom tool. When a program succeeds, the agent saves
it as a private tool. Next time, the agent calls it in one step as
coral.private.<name>({ ... }). Agents find custom tools in the same search
as provider tools, and reuse or edit them instead of starting over. A
workspace owner can publish a private tool as a shared tool under
coral.shared, so everyone in the workspace can call it.
To learn about providers, accounts, policy, and workspaces, read Concepts.
For teams and enterprise
To run Coral for a team, see Run a shared server.
Coral Enterprise is coming soon under its own license. It adds organization-wide governance and support: SSO with SAML, SCIM provisioning, advanced RBAC, an immutable audit log, and support SLAs. See Coral Enterprise or book a demo.
Development
To run Coral from a checkout, follow the
development guide. To contribute, read
CONTRIBUTING.md. The
release notes list what changed
in each version.
Community
Ask questions, share ideas, and show what you built in our Slack community or on GitHub issues.
Security
Do not report security issues in public issues or pull requests. See
SECURITY.md.
Coral is ISO/IEC 27001:2022 certified. Each program runs in its own V8 isolate, and Coral logs every approval.
License
Coral is licensed under the Apache License 2.0. See LICENSE.
The ee/ directory holds Enterprise Edition code under its own license,
ee/LICENSE. A plain cargo build at the repository root
builds only code under the Apache License 2.0. NOTICE states
which license covers which files.
Upgrade from Coral 0.x
Coral v1 replaces the SQL interface with TypeScript programs. It starts from fresh state and does not convert what 0.x stored. If Coral finds 0.x state, it tells you at startup. See Upgrade from Coral 0.x.
来源:Hacker News · github.com

