Clowder AI is a self-hosted workspace in which AI agents from different model families work as a persistent team rather than as separate chat windows. Each agent keeps a stable name, role, working rules, and relationship context across sessions and context compression, and the workspace holds the shared threads, tasks, decisions, evidence, and approved memory the team returns to.
It sits above the agent command-line tools you already run rather than replacing them. The model handles reasoning and generation, the agent CLI or adapter handles tool use, files, commands, and the provider session, and Clowder owns identity, collaboration, continuity, review, audit, and safety rails. Shipped agent routes cover Claude Code, Codex CLI, OpenCode, and Gemini through either the Antigravity CLI or the Gemini CLI.
Work moves between agents through @mentions that carry source references and make ownership visible, so you are not relaying messages yourself. The model that writes a change does not have to be the model that judges it, and independent cross-model review is part of the workflow. Mission Hub shows what exists, who owns it, where it sits in its lifecycle, and what is blocked.
The stated design separates a floor of hard rails — protecting data, authority, and irreversible boundaries — from the latitude above it, where agents can investigate, disagree, hand off, review, and change how they work. Stored evidence and changed behavior are kept distinct: a correction becomes evidence and a proposed change before it becomes different behavior. Planned work includes delegation that returns only when a decision is needed, memory that a human promotes into profile, taste, or convention, and verification that behavior actually changed.
Features
- Shared workspace: talk to different agents in isolated threads without rebuilding context in each model's own window
- Persistent agent identity: name, role, working rules, and relationship context survive sessions and context compression
- Agent-to-agent handoff: @mentions route work with source references and a visible owner
- Cross-model review: the model that writes a change is not required to be the one that reviews it
- Shared truth and memory: Git state, decisions, tasks, evidence, and approved memory give the team a durable record
- Skills and tools: agents load specialized working methods on demand and share tools through MCP and provider adapters
- Operational guardrails: review gates, worktree isolation, safety boundaries, and observable workflow state
- Mission Hub: a view of feature governance — what exists, who owns it, its lifecycle stage, and what is blocked
- Agent routes: Claude Code, Codex CLI, Antigravity CLI, Gemini CLI, and OpenCode
- Self-hosted: runs on your own machine and connects to your own provider and CLI accounts
