Mention Workspace is a multi-agent workspace currently in development with Kiro. People and agents from different providers share one room, but only an explicit @handle in a message wakes the matching agent.
What it does (in development):
- Connect Grok, GPT, Claude, local CLI agents, and custom bots into one workspace.
- A message with no @ is saved as a note and wakes nobody.
- A message can @ one or more agents; only those agents reply.
- An agent's reply can @ another agent, forming a short, explicit chain (maximum 4 hops). Mentions inside code fences are ignored.
- There is no @all, and reading the shared record does not give an unmentioned agent the right to speak.
Frankenstein: the first version stitches together agents that use different OpenAI-compatible API keys and base URLs (for example Grok and GPT) in one TypeScript workspace (Vite + React frontend, Hono room service, SQLite, SSE). Later stages add Anthropic, Gemini, local processes such as Codex or Claude Code, HTTP bots, and A2A Agent Cards. Existing multi-agent frameworks tend to broadcast messages or auto-pick the next speaker; this project implements its own mention router so only an explicit @ can start a turn.
How we use Kiro: the design and architecture are already written. We will use Kiro to turn those specs into implementation, build the mention router and adapters, and write tests for the acceptance cases (a message with no mention wakes nobody; @grok only wakes grok; a mention inside a code fence is ignored).
Status: in development for Kiroween 2026. Design is finalized; implementation has not started yet. Repository: https://github.com/wugang12138/mention-workspace
Built With
- kiro
- llm
- multi-agent
- react
- typescript
