Skip to content

Fusion Code Orchestrator

A panel of small, cheap LLMs advises Claude Code over MCP, and the project measures whether that panel can match one frontier model on cost, speed and quality. Claude Code stays the executor: Fusion returns text (a review, a diagnosis, a plan, a patch to apply) and never edits files or runs anything in your project.

The verdict

The 0.3.0 study: 47 held-out tasks with ground truth, one repeat, neutral judges, $4.42 of live API spend (2026-11-19). The default panel, panel-nano-pick, against Claude Opus 5.5 alone:

Cheaper: yes

About 0.06x Opus's cost per solved task.

Faster: yes

About 0.5x Opus's time per task; 0.2.0's panel was 1.5x slower.

Better: not shown

The quality difference is inside the margin, and one cheap model alone did as well.

Quality against cost per task, one point per arm

The study is small, so "not worse than Opus" is inconclusive, not equal. The numbers, the method, the unedited verdicts and the limitations are in Benchmark results; the same data is in the interactive report, and the run inside a Claude Code session is in Claude Code A/B.

Quickstart

export ANTHROPIC_API_KEY=...   # and/or OPENAI_API_KEY, GOOGLE_API_KEY, OPENROUTER_API_KEY
uvx --python '>=3.11' --from git+https://github.com/alexandre0sheva/fusion-code-orchestrator fusion install claude-code --plugin
claude                         # then try /fusion:review, /fusion:debug or /fusion:ask

Not on PyPI: install from GitHub. Cursor, Codex, the server-only install and troubleshooting are in Integrations.

See what it does

The dashboard's overview: spend against the baseline model

fusion dashboard: spend against the baseline model, runs and benchmarks, read-only (synthetic demo data).

Where to go next

I want to Read
Set it up in Claude Code, Cursor or Codex; see every tool's inputs Integrations
Change models, strategies, budgets or the CLI's behaviour Configuration
Understand the pipeline: lenses, select, escalation, verified patches Architecture
See how cost and price tiers are computed Costs
Know what is sent where, and what is redacted Security
Reproduce or extend the measurements Benchmarking
See what changed in each release Changelog