Models ·
Bring Your Own Model. We Don't Touch Your Tokens, and We Don't Mark Them Up.
TRMNLY's agent runs on Anthropic, OpenAI, Gemini, OpenRouter, any OpenAI-compatible server or a local model through Ollama. With your own key, a turn costs no TRMNLY credits.

You shouldn't have to pay a middleman for the model your agent runs on. TRMNLY's agent runs on the key you bring, with the provider you pick.
Pick your provider
Paste a key in Settings → API Keys for Anthropic, OpenAI or Google Gemini. Three more options need nothing there:1
- OpenRouter. Paste its key in the Providers sheet. Any
vendor/modelid works, whether it's listed in the picker or not, because a/in a model id always means OpenRouter.1,2 OpenRouter offers hundreds of models through one API.3 - Ollama, LM Studio and other local servers. No key needed. Point the endpoint at an Ollama host, or add
/v1for any OpenAI-compatible server such as LM Studio, vLLM or llama.cpp.1 Ollama runs models on your own machine, and its local API needs no authentication.4 - Claude Code, Codex or Gemini CLI. Drive a coding CLI you already have installed, billed to the plan you already pay for.1
Your provider keys live in the macOS Keychain and are never written to disk in plaintext. Settings export leaves them out on purpose.1
Each project can use its own model
Set the app-wide default in Settings → AI Agent → Default model, and pin a different one per project in that project's Providers sheet.5 You could keep a local model on a client repo and Claude on everything else. Two tabs in the same project can even run two conversations on two different models.1
No markup
When the agent runs on your key, every request goes straight from your Mac to the provider. TRMNLY never sits in that path and never marks it up.6 Since 0.1.98, a turn on your own key spends no TRMNLY credits either. Before that, those turns were charged twice: once by your provider and again in credits.7
Local models get a smaller request
A local model has far less room than a hosted one. TRMNLY 0.1.89 fixed a bug where the agent sent all of its ~300 tools to a small local model, Ollama quietly dropped most of the prompt, and a simple "Hello" could come back as a stray line of JSON.8 Now:
- A local model gets a core set of 15 tools. You can change it in the project's Tool allowlist.9
- Each request tells Ollama how much context to load, sized to the request and capped at what the model supports.9
- When a chat grows too big, older context is left out of that request, and the agent tells you. Your saved transcript keeps everything.9
- A model with no tool support, like Gemma, runs as a plain chat instead of throwing an error.8
Sources
- 1.TRMNLY docs, "Agent," Picking a provider. trmnly.dev/docs/agent
- 2.TRMNLY changelog, 0.1.41, "Any model id, plus OpenRouter" (August 3, 2026).
- 3.OpenRouter, "Quickstart." openrouter.ai/docs
- 4.Ollama, "API introduction." docs.ollama.com
- 5.TRMNLY changelog, 0.1.78, "Set the model new conversations start on" (September 4, 2026).
- 6.TRMNLY pricing, "What the meter does not count." trmnly.dev/pricing
- 7.TRMNLY changelog, 0.1.98, "Your own API key no longer spends TRMNLY credits" (October 6, 2026).
- 8.TRMNLY changelog, 0.1.89, "A local model answers instead of printing JSON" and "A local chat keeps going when it outgrows the model" (September 19, 2026).
- 9.TRMNLY docs, "Local LLM," How the agent talks to a local model. trmnly.dev/docs/local-llm