A Slack-native AI agent that refuses to live in your DMs

What it does

Open Claude Tag is a self-hosted Slack bot that acts as a shared teammate inside a channel rather than a private assistant in your DMs. It maintains one persistent session per channel, so every team member sees the same context, threads, and memory. The agent is LLM-agnostic via LiteLLM, supports MCP tools, and runs entirely on your own infrastructure.

The interesting bit

The project inverts the usual Slack bot architecture: sessions are keyed on (workspace_id, channel_id) instead of user_id, which forces shared state and multi-user attribution. After each conversation it runs a Letta-style inner loop to curate MEMORY.md, and after complex tasks it auto-writes SKILL.md playbooks—so the agent literally maintains its own documentation.

Key highlights

• One agent per channel with shared context; everyone picks up the same thread mid-conversation.
• LLM-agnostic via LiteLLM: swap providers with an environment variable, or use different models per channel.
• Agent-curated memory and auto-generated skills: it decides what to remember and authors its own playbooks after multi-step tasks.
• MCP-native tool integration scoped per channel via tools.toml.
• Ambient heartbeat cron that proactively surfaces stale threads or deadlines, but stays silent when there is nothing worth saying.

Caveats

• Semantic recall via Mem0 and Discord/Teams support are listed as Phase 2/3 roadmap items, not yet implemented.
• Token budget controls via BUDGET.md appear in the comparison table but receive no detailed explanation in the current documentation.

Verdict

Teams that want a shared, autonomous Slack agent without cloud lock-in or vendor monoculture should look here. If you just need a quick personal LLM bot in Slack, the channel-scoped complexity is overkill.