Managed.
Picasso-operated provider aliases with budget authorization and alias status checks.
Providers
Choose the model.
Hosted, local, subscription, or API key. The harness stays the same, and the route stays visible.
picasso / providersindexed
Providers
Choose the model
Hosted, local, subscription, or API key. The harness stays the same, and the route stays visible.
ManagedPicasso-operated provider aliases with budget authorization and alias status checks.
BYO keyYour provider account, your billing path, your configured limits.
LocalOllama and vLLM routes run on hardware or infrastructure you control.
ProviderModelsGoalAnthropicOpus, Sonnet, Haiku.Yes.OpenAIGPT and reasoning models.Yes.Microsoft FoundryGPT deployments.Yes.
routes: plan_generation: anthropic/claude-sonnet-4.6 code_generation: qwen/qwen3-coder-nextSupported providers.
Picasso can route work across frontier providers, regional providers, aggregation providers, enterprise deployments, and local runtimes. You can use one route for everything or split planning, coding, reflection, and local verification by model.
| Provider | Models | Goal | Managed | Subscription | API key | Local |
|---|---|---|---|---|---|---|
| Anthropic | Opus, Sonnet, Haiku. | Yes. | Yes. | Claude plans. | Yes. | No. |
| OpenAI | GPT and reasoning models. | Yes. | Yes. | ChatGPT plans. | Yes. | No. |
| Microsoft Foundry | GPT deployments. | Yes. | Yes. | No. | Yes. | No. |
| Gemini Pro and Flash families. | Yes. | Yes. | No. | Yes. | No. | |
| xAI | Grok families. | Yes. | Yes. | No. | Yes. | No. |
| Qwen | Qwen3 and coder families. | Yes. | Yes. | No. | Yes. | No. |
| Xiaomi MiMo | MiMo V2 families. | Yes. | Yes. | No. | Yes. | No. |
| Kimi | Kimi K2 families. | Yes. | Yes. | No. | Yes. | No. |
| DeepSeek | DeepSeek V4 families. | Yes. | Yes. | No. | Yes. | No. |
| Zhipu GLM | GLM-5 and GLM-4.7 families. | Yes. | Yes. | No. | Yes. | No. |
| MiniMax | MiniMax M2 families. | Yes. | Yes. | No. | Yes. | No. |
| OpenRouter | Open model catalog. | Yes. | Yes. | No. | Yes. | No. |
| Ollama | Installed local models. | Yes. | No. | No. | No. | Yes. |
| vLLM | Self-hosted models. | Yes. | No. | No. | No. | Yes. |
Credential modes are explicit.
Subscription passthrough is scoped to supported Anthropic and OpenAI accounts. Other hosted providers are configured through managed or API-key modes. Ollama and vLLM are local-only routes.
BYO key.
Your provider account, your billing path, your configured limits.
Local.
Ollama and vLLM routes run on hardware or infrastructure you control.
Route by task.
Provider policy is part of the product surface. Picasso can use one model for planning, another for implementation, and a local model for reflection when that fits the work.
model accessfour paths
Managedno setupsponsor-supported
Claude / ChatGPTsubscriptionpassthrough
Provider keysBYOdirect billing
LocalOllama / vLLMyour machine
routes: plan_generation: anthropic/claude-sonnet-4.6 code_generation: qwen/qwen3-coder-next reflection: ollama/qwen2.5-coderfallback_chains: default: [anthropic, openai, qwen, kimi, deepseek]Spend with limits.
Set caps per session, day, provider, model, and tool class. The TUI keeps budget state visible, and `/goal` stops or asks for approval when a boundary is reached.
Fallbacks without rewriting the task.
Provider errors can move to the next configured route while the plan, memory, tools, and session state stay intact. The canvas records which model handled each stage.
The catalog is checked, not guessed.
Provider validation jobs, model catalog drift, pricing drift, and cache capability checks make provider choice operational. When a model disappears, a price changes, or a cache hint no longer fits the provider, the admin surface can show the evidence.
Validation jobs.
Manual and scheduled checks record pass, fail, stale, and missing-runner states.
Catalog drift.
Model additions, removals, disabled states, price changes, and capability changes become findings.
Cache policy.
Unsupported cache modes and blocked cache hints can be caught before they become silent routing mistakes.
Provider keys have lifecycle controls.
The provider vault is built around staged keys, disabled keys, rotation review, emergency disable state, alias health, and break-glass read-denial evidence.
Staged before use.
A key can be validated before promotion, and expired validation blocks promotion.
Rotation visible.
Reminder jobs, review tasks, and post-rotation evidence keep key work out of chat history.
Denial evidence.
Break-glass read-denial paths leave operator-visible proof instead of a silent refusal.
Local is first-class.
Ollama and vLLM are normal targets for private work, offline experiments, and hardware-bound development. Local routes can participate in the same plans, goals, subagents, and replay as hosted routes.