IDE Agent Kit: multi-agent workflows for AI coding assistants

[ThinkOffApp]

We released IDE Agent Kit, an open-source Node.js toolkit that connects IDE-based AI agents into team workflows with realtime multi-agent communication.

This could be interesting for the Khoj community since Khoj already acts as a personal AI assistant. IDE Agent Kit provides the infrastructure for multiple AI assistants to coordinate: receiving GitHub webhooks, communicating through shared chat rooms, and executing commands in tmux sessions.

We tested it with three concurrent agents from different providers (Claude, GPT, Gemini) on separate machines, all communicating through shared Ant Farm rooms with sub-10-second response times. No VPN or shared filesystem needed.

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

[aniruddhaadak80]

The interesting part here is the combination of multi-agent coordination and provider diversity without forcing everything into one vendor surface. That can be genuinely useful for coding workflows where search, planning, and implementation benefit from different strengths.

For Khoj specifically, I would be curious how well the kit works when retrieval is treated as a first-class tool rather than just another background capability. That seems like the point where the integration could become more than a generic orchestration layer.

[ThinkOffApp]

Good question about retrieval. The kit treats every agent capability (retrieval, code generation, planning) as something the agent brings to the room, not something the kit manages.

In practice with Khoj: the Khoj agent would join a room with its retrieval capabilities intact. When another agent asks "what does the codebase say about X," Khoj searches its index and posts the answer to the room. The kit handles the communication. Khoj handles the retrieval.

This works better than baking retrieval into the orchestration layer because different agents have different retrieval strengths. Khoj has deep document search. A code agent has AST-level understanding. A web agent has live search. The room lets them combine these without any single orchestrator deciding which retrieval tool to call.