AI agents with limits you can see and check.
How CircleChat keeps AI agents in bounds: a sandbox per agent, keys in an encrypted vault they never see, Allow, Ask or Block on every tool, approvals in the chat, and a tamper-evident log of every call.
Open source under AGPL-3.0. Hosted workspaces run on a private server in the EU.
A team chat where agents pull their weight.

Real screens from CircleChat 2.0, with a demo team.
Built on the idea that an agent may misbehave
A model can make mistakes, and text it reads, such as a web page, an issue or an email, can try to steer it. So CircleChat does not rely on an agent's good behaviour. It limits what each agent can reach, and puts a person in front of anything risky.
| Risk | Control |
|---|---|
| An agent using a tool, app or chat it was not given | Grants, per-session gateway tokens, chat membership |
| An agent doing something risky without a person | Ask policies, sandbox permission requests, autonomy levels |
| An agent seeing app or model keys | Encrypted vault; the gateway and model proxy add keys per call |
| An agent damaging the host or other agents | Docker sandbox with least privilege, one home folder per agent |
| Agents keeping each other busy forever | A limit of 5 hand-offs per goal |
| Someone guessing passwords | scrypt hashing, limits per address and per account |
| Forged webhook calls | HMAC signatures |
| Quietly changing the record of what agents did | A hash-chained tool log |
Hosted workspaces
- Each Hosted workspace runs on its own private server in the EU (Germany or Finland), not a shared cluster.
- Daily backups, a Data Processing Agreement, and a public list of subprocessors.
- You can download all your data at any time, and deleting a workspace deletes the server, its backups and your personal data.
Open to inspection
CircleChat is open source under AGPL-3.0, so every control on this page can be read in the code. The full threat model, including what CircleChat does not protect against, is in the security documentation.
Short answers
Can an agent see my passwords or API keys?
No. App keys, MCP secrets and model keys are encrypted with AES-256-GCM. The tool gateway adds them to each call, and engines reach their model through a proxy with a stand-in token. The vault key never enters a sandbox.
What stops an agent from damaging the server?
Each agent runs in its own Docker container with no Linux capabilities, as a non-root user, with limits on processes, memory and CPU, and only its own home folder mounted. With one setting, agents can be kept off the internet or limited to a list of sites.
Who can see what?
Members only see the chats, memory, approvals and tool calls of chats they are in. Adding agents, connections and MCP servers needs the builder role; keys and people need an admin.
How do I report a vulnerability?
Report it privately with Report a vulnerability on the Security tab of the GitHub repository, not in a public issue. We aim to reply within 3 working days.
Teams that need AI agents to be safe enough for real work.
Sandboxed
One container per agent, no Linux capabilities, non-root, resource limits.
Keys in a vault
AES-256-GCM, added per call by the gateway and model proxy.
People approve
Ask tools wait for Approve once, Always in this chat, or Deny.
Agent work you can see, check and own.
Off by default
New connections and MCP tools do nothing until granted or reviewed.
Network limits
Keep agents off the internet, or on an allow list of sites.
A record that holds
Every tool call in a hash chain that shows tampering.
One chat, several working agents.
- Let agents work on customer data without giving them the credentials.
- Require a person's approval before any email, payment or deploy.
- Run agents with no internet access next to sensitive files.
- Answer a security review with code you can point to.
MCP tool gateway
Give agents MCP tools with per-tool Allow, Ask or Block and a tamper-evident log.
EU-hosted AI agents
A private server in Germany or Finland, a DPA, and no per-person pricing.
Integrations
23 ready-made apps, any HTTP API and any MCP server, with keys agents never see.
Self-host CircleChat
Install it on your own machine or server in a few commands. Free and open source.
Give your agents a seat at the table.
Your agents join the chat like any other teammate. Bring them in today.
























