Perch Roost
Every model. Zero switching.
Roost is the routing engine underneath Perch. It matches each task to a model that is good at it, races providers so a slow one never blocks you, and gives every model the ability to work with images, even the ones that cannot see. You do the work. Roost handles the lineup.
How it works
Matches the task to the model
No single model is best at everything. Roost reads what you are asking for and sends it to one that is suited to it, so you get a good result without shopping for the right name yourself.
Races providers for the answer
Roost can run more than one provider at once and take the first good answer. When a provider is slow or down, another one responds. A single bad provider never leaves you stuck waiting.
Keeps one model on the job
The model chosen at the start of a task stays on it to the end. Work does not drift or change voice halfway through because something switched underneath you.
Gives every model the ability to see
Some of the strongest models are text-only and cannot open an image at all. Perch reads the image for them, so a screenshot, a scanned invoice, or a photo works with any model you are running. On every plan, including the free one, at no extra charge.
Why it matters
You are not betting on one lab.
Most AI tools tie you to a single vendor’s models. When that vendor has a bad quarter, so do you. Perch is model-agnostic by design, which means a new model reaching the frontier shows up in your lineup instead of in a competitor’s product, and the best model for the job is always one you already have.
It also means we can test the claim rather than assert it. We ran nine models through the same build task and published the whole leaderboard, open-weight models included. An open model won one of the two tasks outright.
The lineup
Roost routes across a registry that grows as models ship. A sample of what Pro reaches today:
- GPT-6 Sol
- GPT-6 Luna
- GPT-5.6 Sol
- GPT-5.6 Terra
- GPT-5.6 Luna
- GLM 5.3
- Kimi K3
- Grok 4.6
- Qwen 3.8 Flash
- GLM 5.3 Flash
- GLM 5.2
- DeepSeek V4 Flash 0731 (262K context)
- DeepSeek V4.1 Flash
- Kimi K2.7 Code
- Kimi K2.6
- MiniMax M3
- Qwen 3.8 27B
- Qwen 3.7 Plus
- Nemotron Ultra
- Nemotron 3.5 Lightning
- Qwen 3.6
- DeepSeek V4 Pro
- Inkling
The registry changes as we add and retire models. See the full list and published rates for what each plan includes.
When you want the wheel
Automatic is the right default for most work. It is not the only mode.
Pin a model by hand
On Pro, name any model in the registry and Perch runs it. Pin it in Desktop or the CLI with /model, then drop back to automatic whenever you want.
Four routing tiers
Starter runs the Standard and Standard Max tiers. Pro adds Pro and Pro Max, so heavier work reaches the stronger end of the registry automatically.
One model checking another
In Perch Desktop you can set which model the helpers Perch spawns run on. That is how you have one model check another’s work instead of grading its own. Sub-agents.
Or bring your own
Run Perch on your own cloud key, on any plan, or from the CLI on a model running on your own machine. Your inference, your costs, and no Perch usage limits on it.
What it costs
Metered at list. No markup.
Every model bills against its provider’s published rate, and Perch does not add a margin on top. Starter includes up to $20 of usage each month. Pro includes up to $150, which covers fair-use Roost routing plus up to $75 of models you pin yourself.
Stop shopping for models.
Roost runs on every plan, across Web, Desktop, and CLI, under one account.