Providers.

Curtis.ai ships with 53 built-in providers and supports any OpenAI-compatible endpoint as a custom provider. This doc covers what's built-in, how each one authenticates, per-provider setup quirks, agent compatibility, and the advanced request parameters each card exposes.

Built-in providers

Provider	Auth	Auto-discovery	Agent (tools)	Notes
Anthropic Claude	x-api-key	No	✅	Native Anthropic message format (auto-converted). Agent via native tool use. Vision on Opus/Sonnet/Haiku.
OpenAI	Bearer	Yes (/v1/models)	✅	GPT-6 family at 1.05M context. Vision on most models.
ChatGPT (Sign in)	OAuth	Yes (/v1/models)	✅	ChatGPT Plus/Pro plan billing via OpenAI's official sign-in — no API key. Speaks the Responses API; desktop sign-in, tokens renew automatically. See Subscription plans.
Google Gemini	Bearer	Yes (/v1beta/models)	✅	Gemini's OpenAI-compat endpoint. Vision on Pro/Flash.
Z.ai Coding Plan	Bearer	Yes	✅	Coding plan endpoint — text-only on the coding endpoint even for vision models. For pay-as-you-go use the Z.ai GLM provider instead.
Ollama (Local)	None	Yes (/api/tags)	✅	Native /api/chat dialect. Set a custom endpoint if your Ollama lives elsewhere. Agent needs a tool-capable model.
LM Studio (Local)	None	Yes	✅	Same as Ollama — custom endpoint supported. Agent needs a tool-capable model.
OpenRouter	Bearer	Yes	✅	400+ models via one key. Model IDs are namespaced (openai/gpt-5, anthropic/claude-...).
Groq	Bearer	Yes	✅	Fastest inference available. GPT-OSS models; Llama tiers moved to enterprise-only.
Together AI	Bearer	Yes	✅	Llama, Qwen, DeepSeek hosted.
Fireworks	Bearer	Yes	✅	Similar to Together.
Mistral	Bearer	Yes	✅	Mistral Large 4, Medium 3.5, Codestral.
DeepSeek	Bearer	Yes	✅	V4 Pro/Flash. Function calling supported.
Cohere	Bearer	Yes	✅	Command A family (A Plus flagship), R7B budget.
Meta Muse	Bearer	Yes	✅	Muse Spark 1.3 / 1.2 / 1.1 — 1.05M context, one price across the family. Vision + tools.
Vercel AI Gateway	Bearer	Yes	✅	One key, 20+ upstream providers, namespaced model IDs.
xAI Grok	Bearer	Yes	✅	Grok 4.7 / 4.6 / 4.3, Grok Build.
Perplexity	Bearer	Yes	✅	Sonar models with web search baked in. Served via the Router endpoint, which rejects unknown request fields.
Novita AI	Bearer	Yes	✅	Discount DeepSeek/Llama hosting.
DeepInfra	Bearer	Yes	✅	Similar to Novita.
Chutes AI	Bearer	Yes	✅	DeepSeek, Qwen Coder.
Replicate	Bearer	Yes	✅	Image-gen roots but supports chat now.
Hugging Face	Bearer	Yes	✅	Inference endpoints — any HF model.
Azure OpenAI	Bearer	No	✅	Requires a full deployment URL — see below.
fal.ai	Key <key>	Yes	✅	Chat routes through fal's OpenRouter passthrough at fal.run. Auth header is Authorization: Key <key>, not Bearer.
Cerebras	Bearer	Yes	✅	Fastest inference after Groq. GPT-OSS 120B, Qwen 3.8 27B.
SambaNova	Bearer	Yes	✅	DeepSeek V3, Llama 3.3.
Requesty	Bearer	Yes	✅	Router across many providers.
NVIDIA NIM	Bearer	Yes	✅	Free-credits hosted catalog (100+ models); production pricing is self-host NIM.
Moonshot Kimi	Bearer	Yes	✅	Kimi K3 at 1M context. Temperature/top_p are pinned server-side and never sent. Keys from platform.kimi.ai and .com are not interchangeable.
Kimi for Coding	Bearer	Yes	✅	Kimi's subscription coding plan at api.kimi.com/coding/v1. Plan keys from kimi.com — not interchangeable with platform keys. Quota is calls per 5-hour window.
MiniMax	Bearer	Yes	✅	M3 at 1M context, multimodal. reasoning_effort is honored by M3.1-Flash-Preview only.
Alibaba Qwen	Bearer	Yes	✅	International (Singapore) DashScope compatible-mode. Keys are region-locked. Reasoning via enable_thinking.
Alibaba Coding Plan	Bearer	Yes	✅	Fixed-monthly Model Studio plan. Plan keys are served by the dedicated coding-intl host — standard DashScope keys are rejected there (and vice versa).
AI21 Jamba	Bearer	Yes	✅	Jamba Mini/Large, 256K context. Endpoint requires the /studio prefix. Max tokens hard-capped at 4096.
Upstage Solar	Bearer	Yes	✅	Solar Pro/Mini 4 at 512K context, reasoning + tools.
Nebius	Bearer	Yes	✅	Token Factory (ex AI Studio) — open-weight hosting; full vLLM parameter set.
Baseten	Bearer	Yes	✅	Managed Model APIs — DeepSeek/Kimi/GLM/GPT-OSS at 1M context where offered.
Inference.net	Bearer	Yes	✅	Open-weight catalog plus proxied first-party models; per-token pricing published on /v1/models.
OVHcloud AI Endpoints	Bearer	Yes	✅	European hosting (EU coverage). Catalog prices are EUR; anonymous tier is 2 req/min.
FriendliAI	Bearer	Yes	✅	Serverless open-weight hosting; richest sampling dialect (top_k/min_p/repetition_penalty all real).
GMI Cloud	Bearer	Yes	✅	GPU-cloud inference; contexts/prices live in their console, discovery fills them.
StepFun	Bearer	Yes	✅	Step 5/3.x models, multimodal. This is the international endpoint (USD); CN keys don't work here.
Tencent Hunyuan	Bearer	Yes	✅	TokenHub international — activate models in the console first or calls return 402.
Z.ai GLM	Bearer	Yes	✅	Pay-as-you-go platform endpoint (api.z.ai) — separate key and billing from the Coding Plan. Temperature is 0–1; GLM-5.3 cannot disable thinking. The CN platform (open.bigmodel.cn) speaks the same dialect via a custom endpoint.
BytePlus ModelArk	Bearer	Yes	✅	ByteDance international (Doubao's Seed family, plus hosted DeepSeek/GLM). Keys are region-scoped; model ids carry date suffixes and retire with named replacements. CN region: ark.cn-beijing.volces.com via custom endpoint.
Baidu Qianfan	Bearer	Yes	✅	ERNIE family plus hosted open weights. /v2/models is the richest discovery of any built-in — context, pricing, and features per model. CN platform: real-name verification and CNY billing.
SiliconFlow	Bearer	Yes	✅	International open-weight aggregator (DeepSeek/GLM/Kimi/Qwen at 1M context). The CN site (api.siliconflow.cn) is a separate account/key universe via custom endpoint. enable_thinking is always sent — it defaults to on server-side.
Crusoe Cloud	Bearer	Yes	✅	GPU-cloud inference at aggressive prices. Catalog rotates hard — discovery is authoritative.
Poe	Bearer	Yes	✅	One key, every frontier bot (GPT, Claude, Gemini, Grok, Kimi, …) billed in Poe compute points — prices show 0 because it's point-metered. Unsupported fields are ignored, not rejected.
Featherless	Bearer	Yes	✅	22k+ open-weight models and the richest documented sampling set (top_k/min_p/repetition_penalty all real). Prepaid subscription credits ($50/mo Developer floor); listed prices are the credit rates.
Scaleway	Bearer	Yes	✅	European hosting (GDPR angle). Per-token prices live in the console cost estimator, so cost estimates show as unknown. frequency_penalty is unsupported and dropped.
Reka	Bearer	Yes	✅	Small first-party catalog — Reka Flash 3 has no tool calling, so agent mode needs one of the hosted partner models discovery picks up. Prepaid credits; /v1/models publishes per-model pricing.

Agent compatibility legend

• ✅ — works with the Curtis Agent. Anthropic uses its native tool-use API; everything else speaks OpenAI-style function calling (Gemini via its OpenAI-compatible endpoint, Ollama via its native /api/chat).

Setting up a provider

1. Settings → Curtis.ai → Provider Configuration
2. Find the provider, toggle Enable
3. Paste your API key — it's stored in your OS keychain on Obsidian 1.13+
4. (Optional) Click the refresh icon to auto-discover available models
5. (Optional) Click the crosshair icon to test the connection
6. Pick a default model from the dropdown

The provider now appears in the chat header's model picker.

Advanced request parameters

Every provider card (built-in and custom) has a collapsed Advanced request parameters group. Everything in it is optional — an empty field falls back to the global Generation settings, and an untouched card sends nothing extra. Scope an override to the provider default or to a single model: per-model wins over per-provider, which wins over the global Generation settings.

Controls: Temperature, Max tokens, Top P, Top K, Min P, Seed, Stop sequences (comma-separated), Frequency/Presence/Repetition penalty, an Omit temperature toggle, a Reasoning-effort selector, and Extra body JSON. Omit temperature drops temperature from the request entirely — required for models that reject sampling params (OpenAI reasoning models, Claude released after Opus 4.6). Reasoning effort (off/minimal/low/medium/high/max) is mapped to each provider's dialect: reasoning_effort for OpenAI-style APIs, thinking.budget_tokens for Anthropic, a thinking on/off toggle for DeepSeek and Z.ai, think for Ollama.

• Capability gating — a per-provider capability matrix drops any field that provider's API doesn't accept before the request is sent, so strict APIs (OpenAI, Azure, the Perplexity Router, Fireworks) never see an unknown field. Unsupported controls render disabled in Settings.
• Ollama hardware knobs — Context window, GPU layers, CPU threads, and Keep-alive ride the native /api/chat body as options.num_ctx / options.num_gpu / options.num_thread and top-level keep_alive. Ollama's OpenAI-compatible endpoint can't express them, which is why Curtis speaks the native dialect.
• Extra body JSON — deep-merged into the request body last. Reserved keys: model, messages, stream, stream_options. Everything else is settable — the escape hatch for provider-specific fields Curtis doesn't structure (thinking toggles, reasoning_effort, routing, LM Studio's ttl).
• Copy last request JSON — a debug button on each provider card that copies the exact request body of the last send.

Anthropic Claude

Extended thinking

The Anthropic card's Extended thinking toggle has Claude reason before answering. Reasoning streams into a collapsible "Thinking" block above the reply and never mixes into the answer text; the block collapses once the answer arrives. Thinking budget caps the reasoning in tokens (floor 1024). The budget must sit below the request's max tokens — Curtis raises the cap automatically when it doesn't leave room for an answer.

Supported on Claude 3.7 and the Claude 4 family; for unrecognized model ids (custom endpoints, new releases) the toggle is still honored, and a model that rejects it surfaces the API's own error message. In agent mode, thinking blocks are replayed with each tool result as the API requires. A per-provider or per-model Reasoning-effort value (above) takes precedence over the global toggle, and Extra body JSON wins over both.

Subscription plans and sign-in

Some vendors sell subscriptions whose allowance can be used outside their own apps; others explicitly forbid it. Curtis supports the ones that allow it and deliberately not the rest — the cost of a violation is the user's suspended account, not a broken feature.

Works with your subscription:

• ChatGPT Plus/Pro — Sign in with ChatGPT. The ChatGPT (Sign in) provider uses OpenAI's official OAuth flow — no API key, no client secret. Press Sign in on the provider card, authorize Curtis in the browser window that opens, and requests bill against your plan allowance instead of API credits. Access tokens last about an hour and renew automatically; tokens live in your OS keychain. Desktop only (the sign-in runs a local loopback listener — mobile hides the provider). Requests speak OpenAI's Responses API, and the model catalog is the plan's codex-flavored lineup, which differs from the API-key catalog. OpenAI currently limits plan usage to open-source and local apps; Curtis is no longer open source, so treat plan sign-in as best-effort — the API-key path is unaffected.
• Z.ai Coding Plan — the built-in Z.ai Coding Plan provider is the coding-plan endpoint; paste the plan key from your Z.ai dashboard. The separate Z.ai GLM provider is the pay-as-you-go platform — a subscription plan key will not bill there.
• Poe subscription — the Poe provider bills your Poe compute points through an ordinary API key (create one at poe.com/api/keys); per-token rates match the underlying providers, so a subscriber's points just get spent.
• Kimi for Coding — use the Kimi for Coding provider with a plan key from kimi.com. Plan keys and Moonshot platform keys are not interchangeable.
• Alibaba Coding Plan — use the Alibaba Coding Plan provider with a plan key from the Model Studio console. Plan keys are served by a dedicated endpoint; standard DashScope keys are rejected there.
• MiniMax coding plans — no separate provider: MiniMax plan keys work on the standard MiniMax endpoint.
• Perplexity Pro / Chutes subscriptions — these include monthly API credits delivered as ordinary API keys; paste them into the existing provider.
• Google AI Studio — not a subscription at all: AI Studio issues standard Gemini API keys with a free tier. Paste one into the existing Google Gemini provider — no OAuth involved, nothing to sign into.

Does not work — by design:

• Claude Pro/Max. Anthropic blocks subscription access from third-party apps outright — the tokens are refused at the API level, and Anthropic has suspended accounts that route them through other tools. Curtis uses standard Anthropic API keys only. Pasting a Claude Code OAuth token anywhere will not work and puts your account at risk.
• GitHub Copilot — licensed for first-party clients only.
• Gemini CLI / Code Assist OAuth. Google states that routing the Gemini CLI's OAuth login through third-party tools is a policy-violating use case subject to abuse detection, and has flagged exactly that pattern in third-party apps. The free path that is sanctioned is an AI Studio API key (above) — free tier included, no Google subscription or OAuth required.

Local providers (Ollama, LM Studio)

No API key required. Just enable the provider and ensure the server is running:

# Ollama
ollama serve                              # starts on localhost:11434
ollama pull qwen2.5:7b-instruct           # grab a model

# LM Studio — start the local server from the app UI (default :1234)

Azure OpenAI

Azure uses a per-deployment URL, not a global endpoint. The plugin can't auto-discover models — you must supply the full URL:

https://<resource>.openai.azure.com/openai/deployments/<deployment-name>/chat/completions?api-version=2024-10-21

1. Enable Azure OpenAI in provider config
2. Paste the deployment URL into Deployment URL
3. Paste your Azure API key into the key field
4. Add models manually if needed (no auto-discovery)

Custom providers

Any OpenAI-compatible endpoint works as a custom provider. Common use cases:

• LiteLLM proxy — one local server routing to many providers
• llama.cpp server — self-hosted, fully offline
• Portkey / Helicone — observability gateways in front of OpenAI/Anthropic
• Corporate gateways — internal OpenAI-compat proxies

Adding a custom provider

1. Settings → Custom Providers → + Add
2. Fill in name, endpoint URL, auth type (bearer or none), API key
3. Optionally set a default model
4. Auto-discovery runs if the endpoint exposes /v1/models

Custom providers are agent-compatible as long as the upstream server implements OpenAI function-calling.

Switching providers mid-conversation

Click the model picker button at the top of the chat. The picker shows every enabled provider and their models. Switching is instant — the next message uses the new model. The conversation continues seamlessly.

Per-conversation provider tracking

Every assistant message records which provider and model produced it. Hover an assistant message to see the model name in the meta row above the bubble.

Troubleshooting

"No AI provider configured" — enable at least one provider in settings and ensure it has a valid API key.

"Auth failed" — the API key is wrong, expired, or lacks the required scope. Re-paste it in settings.

"ChatGPT session expired — sign in again" — the refresh token timed out (roughly a month of no use) or was revoked. Press Sign in on the ChatGPT (Sign in) card once and everything resumes.

"This model rejected the image" — you sent an image to a non-vision model. Switch to a vision-capable model via the picker (look for 👁).

"Provider is having issues (5xx)" — the provider is down. Try again or switch providers.

"Tool calls not working" — the agent works with all providers, but the model must support tool calling. On Ollama/LM Studio, pull a tool-capable model (e.g. qwen2.5, llama3.1); models without tool support will ignore the tools silently.

Discovery returns no models — the /models endpoint may require a different auth header or path. File an issue with the provider name + endpoint.