Runner - A local-first agent orchestrator with collaboration mechanism builtin
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
Claude model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
Runner is a local-first AI orchestrator designed to enhance work efficiency by enabling multiple agents to collaborate on a single project. It addresses the gap in inter-agent communication across different providers, leveraging the unique strengths of models like Claude and Codex. For instance, Claude excels in design tasks using pen.dev, while Codex GPT-5.6 with fast mode is noted for coding efficiency. The project is open-source under the GPL-3.0-only license and supports macOS on Apple Silicon.
Hi everyone. Recently, I built a local AI orchestrator to increase my own work efficiency. The idea is simple, I want multiple agents to work on a single project from different angles, with a mechanism to talk to each other. I know Claude Code has already built Claude teams and subagents, but talking between multiple agent providers' models is still an empty space, and I found that Codex and Claude each have their own strength in some particular tasks.
For example, I found Claude extremely powerful when doing design works like design using pen.dev , Codex is not even close at this job. Also, Fable 5 is very good at writing specs and plans, but when it comes to coding efficiency, Codex GPT-5.6 with fast mode on seems to be the most reliable and the fastest option for people with only personal subscriptions.
- Multiple runtime support — Claude Code and Codex in the same crew, talking to each other.
- Each agent runs as a real PTY, so its own TUI is the interface rather than a chat wrapper.
- MCP self-drive — Runner is itself an MCP server, so an agent you are already talking to can assemble a crew and start a mission without you writing the mission description yourself.
- Agent templates and crew templates — predefine roles and working conventions for a group of agents.
- UI: gpui-ce 0.3 — Zed's GPUI, GPU-rendered through Metal. No webview and no JavaScript.
- Terminals: portable-pty, so each agent runs in a real PTY and its own TUI is the interface.
- Storage: SQLite (rusqlite, bundled) plus an append-only NDJSON event log as the source of truth for a mission.
- MCP: rmcp 1.7 in server mode — this is what makes the self-drive feature above work.