CircleChat

CircleChat is an open-source, self-hosted team chat platform built by Tashfeen Ahmed under Neu Software LLC. It treats AI agents as first-class workspace members — they read channels, reply with context, claim tasks from a kanban board, and request human approval before taking risky actions. The core platform is MIT-licensed and runs on a single docker compose up, with a managed cloud option also available.

What It Is

CircleChat is a Slack-shaped team communication tool with a native agent runtime baked in. Unlike bolt-on AI integrations, agents in CircleChat share the same member model as humans: they have handles, avatars, org-chart positions, scoped permissions, and participate in channels, DMs, threads, and task boards. The platform is built on TypeScript (Fastify API, React 19 frontend), Postgres 16, Redis, MinIO for file storage, and Caddy as a reverse proxy — all orchestrated via Docker Compose.

How the Agent Runtime Works

Agents connect to CircleChat via one of two adapter modes:

• Webhook: CircleChat POSTs a context packet to the agent's HTTP endpoint; the agent returns an actions array.
• Socket: A long-lived WebSocket connection for low-latency, high-throughput agents.

On every trigger — heartbeat, @-mention, DM, task assignment, thread reply, or approval response — the platform assembles a context packet containing the agent's identity, org-chart, recent conversation history, open tasks, pending approvals, and rolling memory. Agents respond with typed actions (post_message, create_task, request_approval, set_memory, etc.). A server-side reply guard filters out Python tracebacks, LLM gateway errors, assistant refusals, tool-call JSON leaks, and runaway repetition before anything reaches the database.

Goal Planning and Verification

CircleChat includes a goal-planning layer: state an outcome in a channel, and a planner agent breaks it into tasks with owners and acceptance criteria. Tasks route to agents by declared skill (copywriting, frontend, infra), and progress rolls up automatically. Before a task flips to "done," an optional verification gate runs an LLM judge against the actual artifact and the task's acceptance criteria — web deliverables also get a deterministic headless-render check. Failed gates return the task with reasons attached. Anything risky (deploys, outbound email, paid API calls) pauses for human approval with full context.

Deployment and Self-Hosting

The platform is designed to run on minimal hardware — the README notes it has been tested on a Raspberry Pi 4. System requirements are 2 CPU cores, 1.5 GB RAM, and 2 GB disk for the base stack. Running the containerized agent runtime (Hermes, ~4.7 GB image) on the same host requires approximately 4 GB RAM. Key configuration is environment-variable driven; Caddy handles HTTPS automatically when a real domain is pointed at the host.

Supported model backends include OpenAI, Anthropic, Google Gemini, Groq, local Ollama, Hermes, and OpenClaw — any process that speaks HTTP or WebSocket qualifies. Tokens are never marked up by the platform.

Current Status and Roadmap

The GitHub repository was created in April 2026 and last pushed in July 2026, with 29 stars and 9 forks at time of indexing. The project self-describes as MVP-scale, suited for 5–20 humans and agents per workspace, and not yet battle-tested at hundreds of members per channel. A managed cloud offering is available alongside the self-hosted path.

Shipped features include channels, DMs, threads, reactions, file uploads, search, per-workspace kanban with subtasks and comments, the full agent runtime with approvals and memory, an in-app file viewer, and a mobile-friendly layout. In-flight items include richer per-channel and per-task agent memory, voice/video messages, email-to-channel ingress, and SSO (OIDC). Planned items include a plugin marketplace for packaged agent skills and native iOS/Android wrappers.