AI providers
The assistant, the summarizer and every AI feature run on one active provider chosen in Settings → AI Provider. Each provider has a main model and a summarizer model, its own authentication, and its own usage display. Switch providers whenever you like; conversations and boards are unaffected.
Providers come in two kinds:
- Built in - always available, no setup beyond a key: DeepState AI, OpenAI-compatible, OpenRouter and Ollama.
- Plugins - Claude, OpenAI and OpenCode. They are open source and turned on one at a time from Settings → AI Provider → Plugins. See Provider plugins below.
| Provider | Kind | Sign in with | Cost to you | Vision | Notes |
|---|---|---|---|---|---|
| DeepState AI | built in | your license | included allowance | ✓ | Zero setup; capped monthly budget |
| OpenAI-compatible | built in | API key + base URL | your provider's bill | model-dependent | DeepSeek, Kimi, GLM, Qwen, MiniMax, NVIDIA, Groq, Mistral, xAI, a local server, … |
| OpenRouter | built in | API key | prepaid credits | model-dependent | Hundreds of models from every lab |
| Ollama | built in | nothing - runs on your machine | free | model-dependent | Fully offline; tool support varies by model |
| Claude (Anthropic) | plugin | API key or Claude Pro/Max subscription | your Anthropic bill or your subscription | ✓ | |
| OpenAI | plugin | API key or ChatGPT Plus/Pro subscription | your OpenAI bill or your subscription | ✓ | Subscription login also unlocks the Codex models |
| OpenCode Zen | plugin | API key | per-token | model-dependent | Curated frontier models |
| OpenCode Go | plugin | API key | $10/month subscription | model-dependent | Budget models |
DeepState AI
Every license includes model access through DeepState's router, so you can research the moment you activate. Select DeepState AI, pick a model and go.
- Models:
anthropic/claude-sonnet-4,anthropic/claude-haiku-4.5,moonshotai/kimi-k2.5, plus any free models we enable. All support tools and vision. - Budget: $2/month on Personal, $8 on Pro, $8 per seat on Enterprise, resetting monthly. Usage appears as a bar in Settings → License; it turns amber near the limit and requests are declined at it until the next cycle.
- How it works: requests go to
router.getdeepstate.app, which forwards them to the model provider under a key that is scoped to your license and budget. The router logs token counts for billing and nothing else; prompts and responses are not stored. See Security & privacy.
When the allowance is not enough, switch to your own provider - there is no cap when you bring your own key or subscription.
OpenAI-compatible
Most model providers now speak the same API as OpenAI (Chat Completions). This provider connects to any of them with a base URL and an API key, which covers the Chinese labs, NVIDIA's hosted models, the fast-inference clouds and servers you run yourself.
Setup: choose OpenAI-compatible, then either pick your provider from the searchable list or paste its base URL, and save your API key. The list comes from models.dev and has about 190 providers.
| Provider | Base URL |
|---|---|
| DeepSeek | https://api.deepseek.com |
| Moonshot (Kimi) | https://api.moonshot.ai/v1 (China: api.moonshot.cn) |
| Z.ai / Zhipu (GLM) | https://api.z.ai/api/paas/v4 (China: open.bigmodel.cn/api/paas/v4) |
| Alibaba Model Studio (Qwen) | https://dashscope-intl.aliyuncs.com/compatible-mode/v1 |
| MiniMax | https://api.minimax.io/v1 |
| StepFun | https://api.stepfun.ai/v1 |
| Volcengine Ark (Doubao) | https://ark.cn-beijing.volces.com/api/v3 |
| SiliconFlow | https://api.siliconflow.com/v1 |
| Xiaomi MiMo | https://api.xiaomimimo.com/v1 |
| NVIDIA (build.nvidia.com) | https://integrate.api.nvidia.com/v1 |
| Groq | https://api.groq.com/openai/v1 |
| Mistral | https://api.mistral.ai/v1 |
| xAI (Grok) | https://api.x.ai/v1 |
| Google Gemini | https://generativelanguage.googleapis.com/v1beta/openai |
| Together, Fireworks, Cerebras, DeepInfra, Hugging Face, … | pick from the list |
| vLLM, LM Studio, llama.cpp, LiteLLM | your server, e.g. http://localhost:8000/v1 |
Models. The model list is read from the endpoint itself ({base URL}/models), so you see
exactly what your key can use. When the URL belongs to a provider on models.dev, each model also
gets its context window, vision and tool-calling support and price from there. If the endpoint
has no model list, the models.dev list for that provider is used instead. After your key is
saved, DeepState chooses a starting pair: a recent flagship for the main model and a cheap recent
model for the summarizer. Change either one whenever you like.
To use a model that isn't listed, type its exact id in the model search and choose
Use model id.
Your key. DeepState checks a key before saving it, and the check costs nothing. It asks the endpoint for its model list, then sends an empty chat request that the provider refuses before running any model. If the key is wrong, you see the provider's own error message.
Good to know
- The assistant works the board through tool calls, so pick a model that supports them. The list shows what models.dev knows; models it doesn't know are assumed to support tools but not images.
- Models that think before answering (DeepSeek, Kimi, GLM, Qwen, …) stream that reasoning into the agent the same way other providers do.
- Local servers that don't check keys still need something in the key field; any text works.
- A few providers need part of the URL filled in, such as a Cloudflare account id. The list
shows these as
${...}; replace that part and press Enter. - Usage shows the tokens used this session, plus an estimated cost when models.dev has prices for the model.
Claude (Anthropic)
Plugin; enable it under Plugins first. Two ways to sign in:
- Claude Pro/Max subscription. Click Login with Claude Pro/Max, approve in the browser window that opens, copy the code shown and paste it back into DeepState. Usage then draws on your subscription's 5-hour and 7-day windows, which the Provider usage section displays with reset times. No API key, no per-token cost.
- API key from console.anthropic.com. Pay per token; the usage section shows tokens and an estimated cost for the session.
Models are fetched live; recent defaults are claude-sonnet-4-5 for the main slot and
claude-haiku-4-5 for the summarizer. Sonnet-class is the recommended default for research;
Opus for hard synthesis; Haiku for the summarizer.
OpenAI
Plugin; enable it under Plugins first.
- ChatGPT Plus/Pro subscription. Click Login with ChatGPT Pro/Plus; DeepState opens the
browser for you to approve and receives the token on a local callback. This route gives access
to the Codex model family (
gpt-5.x-codex) in addition to the standard models, at no per-token cost. - API key from platform.openai.com. Pay per token.
Defaults: gpt-5.2 main, gpt-5-mini summarizer.
OpenRouter
One key, every model. Get a key and add credits at openrouter.ai, paste
it in, and the model list is fetched live with pricing tiers shown. The usage section shows your
remaining credits. Default anthropic/claude-sonnet-4.5 / anthropic/claude-haiku-4.5.
Good for trying a model DeepState doesn't otherwise offer (Gemini, DeepSeek, Llama, Mistral, …) without another account. Models without tool support will chat but cannot work the board.
OpenCode Zen and OpenCode Go
Plugin; enable it under Plugins first. Two model services from OpenCode. Zen offers frontier models (Claude Sonnet 4.5, GPT-5.2, Gemini 3 Pro, Kimi K2.5, GLM-5) per token; Go is a $10/month plan with capable budget models (Kimi K2.5, MiniMax M2.x, GLM-5). Paste the API key from your OpenCode account.
Ollama (local)
Run models on your own hardware with Ollama. Install it, ollama pull a
model, and DeepState discovers what you have at the Ollama server URL (default
http://localhost:11434; change it to point at a machine on your network).
- Nothing leaves your computer. This is the option for air-gapped or sensitive work.
- Tool use is model-dependent. The assistant needs a model that supports function calling to create nodes; Llama 3.x, Qwen 2.5+, Mistral and similar instruct models do. Smaller models may call tools unreliably; turn on Smart tool loading to shorten the prompt.
- Speed and quality follow your GPU. A 7-8B model is workable for summaries; 30B+ for the main assistant if you have the memory.
Provider plugins
Claude, OpenAI and OpenCode ship as provider plugins. They're open source, at github.com/DeepState-Inc/deepstate-plugins, so you can read exactly what each one sends and where.
- Opt-in. Open Settings → AI Provider → Plugins and enable the ones you want. Each one appears in the provider list straight away, without a restart. A plugin that isn't enabled is never loaded. If you used one of these providers before plugins existed, it was enabled for you when you updated.
- Disabling keeps your login. A disabled plugin keeps its saved key or subscription login, so turning it back on needs no new sign-in.
- Limited by design. A plugin is a single JavaScript file. The only libraries it can use are the AI SDK libraries built into DeepState. It cannot read your files or boards, and it cannot start programs. Keys and logins are stored by DeepState, encrypted with your operating system's keyring, and handed only to the plugin they belong to.
- Updates. A copy of each plugin ships with the app. On start, DeepState checks the repository's registry for newer versions. Each downloaded file is checked against the SHA-256 hash published in the registry, and updates take effect on the next restart.
- Write your own. The repository's README documents the plugin API, including authentication by key or browser login, model lists and usage display. It also has a checker and tests. Pull requests for new providers are welcome. For a provider that already speaks the OpenAI API, you don't need a plugin: use OpenAI-compatible.
Choosing models
Each provider has two slots:
- Model - what the assistant reasons and talks with. Reasoning depth (the toggle in the AI panel) applies here.
- Summarizer model - for article summaries, feed relevance, narrative summaries, compaction and the Autoresearch supervisor. Pick something fast and cheap. Details in Summarizer.
The model picker is searchable and shows a pricing tier per model where the provider publishes one. Lists are fetched from the provider (or from models.dev) and cached; a static fallback list is used offline. If a model disappears from its provider, DeepState switches to the closest replacement, and tells you when it was one you picked.
Usage display
Settings → AI Provider → Provider usage adapts to the provider: subscription windows for Claude OAuth, tokens and estimated cost for API keys, credit balance for OpenRouter, a link to the License tab for DeepState AI, session tokens (and cost where known) for OpenAI-compatible, nothing for Ollama. The agent's header shows the model name and the token count of the current turn.
Keys and credentials
Keys and OAuth tokens are stored encrypted with your operating system's keyring. They are sent only to the provider they belong to. Clearing a key or logging out removes it from disk. Nothing about your provider configuration is synced to collaborators or included in board exports or backups - after a restore on a new machine you sign in to providers again.