Claude leads. Your other AIs do the small jobs.
RouteAI lets Claude hand unit tests, scripts, docstrings, build files and bulk edits to the AIs you already have: the Ollama models on your computers, and the free daily quotas of providers like Gemini, Groq or OpenRouter. Free capacity first, paid only if you allow it, and Claude writes the brief, checks the result and keeps the final word.
Python 3.11 and Ollama, nothing else · API providers optional · Windows, macOS, Linux · MIT license
A team of junior developers on your own hardware
One small MCP server with no dependencies, a routing brain that learns, and a skill that tells Claude when delegating is worth it.
Routed by power and by cost
Every task has a category. complex and code go first to the fast GPU machine; tests, scripts, build and docs to the slower ones or to a free provider tier. Free capacity is used before paid, and work overflows when the preferred machines are busy.
It learns which model to trust
The router ranks machines and models by measured quality and speed — including model load time and queue. Scores come from the benchmark and from Claude rating every real result as good, fixed or rejected.
Claude keeps the governance
Claude decides what to delegate, writes a precise brief, runs the tests and reads the diff. Design, security-sensitive code and cross-cutting changes stay with Claude. The benchmark only suggests configuration changes; you approve them.
Fewer tokens, for real
The server reads your project files itself and writes results straight to disk, returning only a short preview. Claude never pastes a file in or reads a long answer back, and every task reports the tokens it saved.
Parallel batches
One instruction, forty files: tests for every module, a docstring on every function, the same migration everywhere. Tasks spread over all your machines at once and land where you tell them, e.g. tests/test_{stem}.py.
Nothing to install but Python
Pure standard library: no pip, no SDK, no Docker. Add an API provider whenever you want — Gemini, Groq, OpenRouter, DeepSeek, Mistral, OpenAI or your own endpoint — with the key in an environment variable, never in a file.
How the pieces fit
Claude Code talks to one small local server; the server talks to the Ollama machines on your network and, if you allow it, to the API providers you added.
- Router: picks the node and model for each task, free before paid
- Learned quality and speed of every model
- Self-grading benchmark and token accounting
- Claude writes a briefA few precise lines and the file paths — never the file contents, so Claude does not pay to read them.
- The server routes itIt reads the files, picks the best free machine and model for the category, and the result is written straight to disk.
- Claude verifies and ratesIt runs the tests, reads the diff and marks the result good, fixed or rejected. Routing learns from every rating.
- You see what you savedEvery finished task reports the Claude tokens it saved and the session total. Tiny tasks can come out negative, and the numbers say so.
Running in five minutes
If Ollama already runs on your machines, you are almost done.
- Install the pluginTwo commands in Claude Code add the marketplace and install RouteAI: the MCP server, the delegation skill and the
/routeai:statusand/routeai:benchcommands. - Describe your machines
/routeai:setupprobes each Ollama server and writes~/.routeai/fleet.toml;/routeai:add-aiadds a provider — its key stays in an environment variable, and you decide whether your files may reach it. - Work as usualWhen a plan contains tests, scripts or boilerplate, Claude delegates them, verifies what comes back and rates it. Ask
/routeai:statusany time to see where work goes and how many tokens were saved.
Install
# in Claude Code /plugin marketplace add bdbais/routeai /plugin install routeai@bais # describe your machines: models are picked and suggested per machine /routeai:setup gpu=http://192.168.1.13:11434 local=http://localhost:11434 # let the fleet learn your hardware, then check routing and savings /routeai:bench /routeai:status
On machines other than your own, start Ollama with OLLAMA_HOST=0.0.0.0 so it listens on your network. The plugin finds Python 3.11+ by itself (py, python3 or python).
A benchmark that grades itself
Run it now and then until the fleet knows your hardware: often while it is still learning, monthly once the scores are stable. Paid providers are left alone unless you ask for them.
Graded, not guessed
Ten real tasks with automatic checks: code runs against hidden unit tests, generated tests must catch deliberately planted bugs, scripts are executed, TOML is parsed, docstrings are compared with the AST.
Squeezes your hardware
It measures tokens per second, model load time, how much of each model fits in VRAM and how throughput grows with parallel requests — then tells you the context size and parallelism that work best.
Recommendations, not surprises
Each run writes a report: the best model per category, models to drop, a faster node that deserves the heavy tier. Claude turns it into a proposed fleet.toml change for you to approve.
Knows when to stop
Every model and category pair becomes stable after five consistent samples. Status tells you when a new run is due, and real work keeps refining the scores in between.
Your code stays on your network
Delegating is only useful if it does not open new holes.
Nothing leaves your machine by default
Local models see your files; a provider you added does not, unless you set send_files for it. Ollama :cloud models, which run on ollama.com, are ignored unless you allow them explicitly.
Confined file access
The server reads and writes only inside the current project and any folder you list. Paths that escape it are refused.
No telemetry
Statistics, logs and reports stay in ~/.routeai on your computer. The plugin talks only to the Ollama servers you configured.
Keys and budgets
API keys live in environment variables: never in the configuration, never in the chat, never in a log, and requests never follow redirects. Each provider carries its price and a daily cap in requests, tokens or dollars; when the cap is reached the work goes back to your own machines.
Free and open source, and it will stay that way
RouteAI is released under the MIT license: use it at work, change it, share it. If it saved you tokens, a small donation helps keep it going.
Questions
Which models should I use?
Setup suggests coder models that fit each machine — on a 12 GB GPU qwen2.5-coder:14b (about 50 tok/s) and the mixture-of-experts qwen3-coder:30b, on a CPU-only laptop 3–7B models — and downloads them only with your approval. For providers you pick from the models they offer. The benchmark then shows what really works.
How many tokens does it save?
Every finished task tells you: the files the worker read plus what it wrote, minus the brief Claude wrote and the result Claude read. Substantial tasks and batches save thousands of tokens; tiny ones can cost more than they save, and the report says so — including what a paid provider cost you.
What happens if a machine is off?
It is marked down and the work goes to the other machines. If none can take a category, the tool says so and Claude simply does the task itself.
Is a laptop without a GPU useful?
Yes, for scripts, tests and docs with small models, and as overflow when the GPU machine is busy. Keep one task at a time on it so your own work stays responsive.
Can I add Gemini, Groq or OpenAI?
Yes: /routeai:add-ai gemini gem1. Any OpenAI-compatible API works, including your own vLLM or LM Studio. The key goes in an environment variable, you decide whether your files may be sent, and you set a daily cap in requests, tokens or dollars.
Is it an Anthropic or Ollama product?
No. It is an independent open-source project, not affiliated with Anthropic or Ollama. It uses Claude Code's public plugin and MCP interfaces and Ollama's public API.