Provider Support Matrix
LARGESTACK separates adapter exists from verified production support.
Use largestack.provider_support_matrix() to inspect capabilities at runtime.
Status meanings:
verified: covered by local tests and known primary implementation path.partial: adapter and routing exist, but exact tool/structured-output behavior depends on the provider/model or live credentials.experimental: useful path, but must be live-tested before production claims.adapter_only: source adapter exists, but should not be marketed until wired and verified.
Public proof tiers:
validated: verified path with local tests and release evidence.compatible: adapter or OpenAI-compatible path exists; require model-specific live E2E before strong claims.experimental: useful but not production-claimable yet.skipped: adapter/source exists, but omit from marketing until wired and validated.
Recommended production claim
Say:
LARGESTACK supports OpenAI, DeepSeek, Anthropic, LiteLLM, Ollama/local models, and many OpenAI-compatible providers through a verified/partial capability matrix.
Do not say every listed provider has equal production-grade tool-calling support until live E2E gates pass.
Public Provider Proof Register
This table is the public claim source. Keep it aligned with
largestack.provider_support_matrix().
| Provider | Public tier | Current proof | Next gate before stronger claim |
|---|---|---|---|
| OpenAI | validated | Primary adapter path and local tests | Fresh live E2E for chosen public model |
| DeepSeek | validated | Live difficult-project evidence plus OpenAI-compatible path | Repeat on release branch with current key |
| Anthropic | adapter-only | Native adapter + tool-mapping tests; not live-verified (no key tested) | Run a live E2E with a real sk-ant- key to reach "validated" |
| LiteLLM | compatible | Gateway integration present | Per-downstream-provider live E2E |
| local OpenAI-compatible | compatible | Generic OpenAI-compatible path | Validate target runtime/model tool behavior |
| Ollama native | verified | Live-verified local chat (qwen2.5/llama3.2); tools via ollama_openai/ |
— |
| Ollama OpenAI-compatible | verified | Live-verified /v1 chat + function-calling locally |
— |
| Google/Gemini | verified | Live-verified chat + tools + structured output | — |
| Groq | compatible | OpenAI-compatible path | Model-specific tool/structured-output E2E |
| Mistral | compatible | OpenAI-compatible path | Model-specific tool/structured-output E2E |
| Cohere | compatible | Chat/embedding adapter present | Tool parsing and streaming proof |
| Bedrock | compatible | Adapter present, AWS credential gated | AWS live E2E and error/fallback proof |
| Azure OpenAI | compatible | Azure OpenAI-compatible path | Deployment-specific live E2E |
| OpenRouter | compatible | Aggregator path present | Routed-model proof for claimed model |
| Perplexity | compatible | Chat/research path present | Tool/structured-output proof if claimed |
| Cerebras | compatible | OpenAI-compatible path | Live model-specific E2E |
| SambaNova | compatible | OpenAI-compatible path | Live model-specific E2E |
| xAI | compatible | OpenAI-compatible path | Live model-specific E2E |
| AI21 | compatible | Adapter present | Live chat/tool capability proof |
| Lepton | compatible | OpenAI-compatible path | Live model-specific E2E |
| NVIDIA | compatible | OpenAI-compatible path | Live model-specific E2E |
| Anyscale | skipped | Adapter source exists | Wire into release gate before listing publicly |
| Cloudflare | skipped | Adapter source exists | Wire into release gate before listing publicly |
| Databricks | skipped | Adapter source exists | Workspace auth/live E2E |
| Replicate | skipped | Adapter source exists | Live chat path and error handling proof |
| Voyage | skipped | Embeddings-oriented adapter | Do not market as chat-provider support |
Local LLM guidance
ollama/<model>: native Ollama chat path. Good for local chat/summarization.local/<model>withLARGESTACK_OPENAI_COMPATIBLE_BASE_URL: generic OpenAI-compatible endpoint, good for vLLM, LM Studio, Ollama/v1, or enterprise gateways.ollama_openai/<model>withLARGESTACK_OLLAMA_OPENAI_COMPAT=true: Ollama OpenAI-compatible endpoint. Tool calling only works if the local runtime/model supports it.
export LARGESTACK_OPENAI_COMPATIBLE_BASE_URL=http://localhost:11434/v1
export LARGESTACK_OPENAI_COMPATIBLE_API_KEY=ollama
from largestack import Agent
agent = Agent(name="local", llm="local/llama3.2")