Skip to content
MCP tool gateway

An MCP gateway that asks before your agents act.

CircleChat's tool gateway gives AI agents MCP tools with per-tool Allow, Ask or Block, keys they never see, review for new tools, and a hash-chained log of every call. Open source and self-hostable.

Open source under AGPL-3.0. Works with every engine CircleChat runs.

How it works

A team chat where agents pull their weight.

A group chat where Juno posts a four step launch plan, saves the launch date to team memory and hands research to Pip.

Real screens from CircleChat 2.0, with a demo team.

Why put a gateway between agents and MCP servers

Giving an agent an MCP server usually means giving it every tool on that server, with the server's credentials, and trusting the model to behave. A model can make mistakes, and text it reads, such as a web page or an issue, can try to steer it.

CircleChat does not rely on the agent's good behaviour. Every agent session gets its own private MCP endpoint that lists only the tools it was granted. The gateway checks each call against your policy, adds the secret, runs it, and writes it to a log.

What the gateway does on every call

  • Checks the session's token. Tokens are random, per session, held in memory and revoked when the session ends.
  • Checks the grant: is this tool given to this workspace, chat or agent?
  • Applies the policy: Allow runs, Ask waits for a person, Block refuses.
  • Adds the key from the vault, so the agent never holds it.
  • Records who asked, as whom, who approved and the cost, in a hash chain that shows tampering.

Works with every engine

Hermes Agent, OpenClaw, Claude Code, opencode, Codex CLI, Gemini CLI and goose all reach the same gateway, so one set of rules covers every agent. Ready-made apps and HTTP APIs go through it too.

Questions

Short answers

What is an MCP tool gateway?

A single place between your agents and their tools. Agents call tools through it, and it decides whether each call may run, adds the credentials, and records what happened. In CircleChat, the gateway serves Model Context Protocol (MCP) tools to every agent engine.

How are new MCP tools handled?

They stay off until a person reviews them. That includes tools an MCP server adds later, so a server cannot quietly give agents new powers.

Can I require approval for some MCP tools?

Yes. Each tool is Allow, Ask or Block. An Ask tool turns into an approval card in the chat with the exact call, and the agent waits for Approve once, Always in this chat, or Deny.

Which MCP servers can I add?

MCP servers over HTTP, and stdio servers that run on the CircleChat server. Secrets for them go in the encrypted vault.

Built for

Teams giving AI agents real tools, who need control over every call.

Private endpoints

Each agent session gets its own MCP endpoint with only its granted tools.

Reviewed tools

New MCP tools stay off until a person turns them on.

Approvals in chat

Ask tools become approval cards where the team already is.

What changes

Agent work you can see, check and own.

Least privilege

Grants per workspace, chat or agent, and nothing by default.

No keys in sandboxes

Secrets live in an AES-256-GCM vault and are added per call.

Proof of what happened

A tamper-evident log of every tool call.

Use cases

One chat, several working agents.

  • Let agents read from an internal MCP server while writes need approval.
  • Give one agent a database tool without giving it to the rest.
  • Review a third-party MCP server's tools before any agent can call them.
  • Show an auditor exactly what an agent did, and who approved it.

Give your agents a seat at the table.

Your agents join the chat like any other teammate. Bring them in today.