Repo of the Day
yatuk/multi-model-orchestra
Published: Nov 21, 2026
Open repository ↗Route coding work across Claude, GPT (Codex), Kimi and DeepSeek: one runner with hard status gates, worktree-only writes, a blind council and a benchmark
Summary
multi-model-orchestra is a set of agent skills and Python/shell tooling for routing coding work among Claude, GPT through Codex CLI, Kimi, and DeepSeek. A host agent acts as the conductor, while a shared runner applies hard status codes, worktree-only writes, timeouts, and read-only safeguards before delegated results are accepted.
What it is useful for
It is useful when an engineer wants to delegate implementation, seek a second opinion, compare decisions across model providers, or determine whether a locally configured coding lane can complete a task reliably. The routing policy distinguishes hard, routine, and bulk work, with Claude tiers running as the host agent’s own subagents and other providers accessed through their CLIs.
The blind council can send the same question to three model families, shuffle their answers into letters, and exclude failed lanes from the vote. This is useful for design reviews or other decisions where the identity of the responder should not influence the initial assessment.
The included benchmark runs bug-fixing, reasoning, and log-labeling tasks against configured lanes. It is intended as a smoke test rather than a leaderboard: the README reports that low-effort GPT Luna failed visible tests and scored 56/60 on bulk labels, while Kimi Highspeed was unavailable because of a plan limit.
The safety boundaries are explicit. Read-write runs must use a linked Git worktree. Read-only Kimi and DeepSeek runs use a throwaway detached worktree because those CLIs may still edit files. Briefs are sent to external services and must not contain secrets or personal data. The unattended otomode feature is described as a guard rail, not a full sandbox.
How engineers can use it
Requirements are Bash, Git, timeout, Python 3, and signed-in CLIs for the providers being used: codex for GPT and kimi for Kimi and DeepSeek. Windows users are directed to Git Bash.
git clone https://github.com/yatuk/multi-model-orchestra
cd multi-model-orchestra
./install.sh --link
~/.claude/skills/orchestra/scripts/mmo-preflight
The documented project-local alternative is ./install.sh --dir PATH. The preflight command runs offline and reports which lanes are available.
Delegate work through the common runner:
mmo-run <codex|kimi|deepseek|claude> --brief FILE --out FILE \
[-m MODEL] [-e EFFORT] [--mode ro|rw] [-C DIR] [-t SECONDS]
Its output begins with a status header. Exit 0 means an answer exists but still requires review; 3 through 7 distinguish failure, timeout, empty output, provider limits, and authentication problems. A failed lane should trigger a fallback rather than being treated as an answer.
For unattended work, the README also documents:
python ~/.claude/skills/orchestra/skills/otomode/scripts/otomode.py install
Finally, bench/run.sh runs the local benchmark, including the documented example:
bench/run.sh codex:gpt-6.1-sol:high kimi:kimi-code/k3.1-pro