Multi-runtime teams

Mix Claude Code, Codex, and Gemini freely inside one team. They all speak MCP over stdio against the same mailbox; the manager doesn’t know or care which CLI a worker happens to be running.

Declaring runtimes

Each runtime lives in runtimes/<name>.yaml:

runtimes/claude-code.yaml

binary: claude
supports_mcp: true
session_resume: "--continue"
default_model: claude-opus-4-7

Reference one from an agent spec:

workers:
  dev1:
    runtime: codex
    model: gpt-5-codex
    reports_to: manager

The agent-wrapper.sh dispatches on $RUNTIME and calls the matching binary with the right flags.

When to pick which

Runtime	Strong at
Claude Code · Opus	planning, orchestrating, long system prompts
Claude Code · Sonnet	fast, cheap tool use; frontend refactors
Codex · GPT-5	deep reasoning on complex backend patches
Gemini · 3.0 Pro	1M-token context for research / large-corpus reads

Cost

Each runtime reports cost differently. teamctl budget aggregates whatever has been recorded in the budget table. Runtime-specific cost parsers are pluggable and land with the runtime adapter itself.

Example

See examples/multi-runtime/ — one Claude-Code manager directs a Codex backend dev and a Gemini researcher.