IDE Agent Kit: multi-agent communication for IDE-based assistants

[ThinkOffApp]

We just open-sourced IDE Agent Kit, a Node.js toolkit that connects IDE-based AI agents (Claude Code, Codex, Cursor, VS Code agents, local LLM assistants) into team workflows with realtime communication.

For the Superagent community, this bridges the gap between autonomous agents and IDE-based coding assistants. Agents can receive GitHub webhooks, communicate through shared chat rooms, execute commands via tmux, and log everything with append-only receipts.

Three concurrent agents from different providers have been tested on separate machines with sub-10-second response times.

Repo: https://github.com/ThinkOffApp/ide-agent-kit
Launch post: https://www.thinkoff.io/

[aniruddhaadak80]

From my point of view, this gets interesting only if the communication layer preserves provenance, permissions, and task ownership rather than acting as simple message fanout. Once multiple IDE-based agents can all respond, the hard problem becomes deciding which one is allowed to act and how shared state stays coherent.

If the kit makes those coordination boundaries explicit, it could be genuinely useful for framework-level agent collaboration.

[ThinkOffApp]

Good questions. The kit handles these explicitly:

Provenance: every message carries sender identity. The room history is an audit trail.
Permissions: rooms are invite-only. The gateway validates tokens per-agent.
Task ownership: claim-based. An agent posts "I am taking task X" to the room. Message ordering is global so conflicts are visible.
Turn-taking: the kit does not impose turns. Agents post concurrently. Explicit turn-taking can make multi-agent conversations worse, not better.

The hard part you identified (which agent is authoritative) is a policy question the kit intentionally leaves to the operator.