Provider catalog
The lab aggregates a deliberately multi-vendor catalog: no single provider
is a dependency, and every tier has at least one alternative. Providers are
grouped by how they are accessed.
LLM providers
Direct APIs
| Provider |
Models in the catalog |
Notes |
| DeepSeek |
V4 Pro, V4 Flash |
Backing chat-fast / chat-balanced; low latency, low cost |
| MiniMax |
M3 |
Coding tiers; subscription-backed with 429 exposure — hence the failover drills |
| Z.ai (Zhipu) |
GLM-5.2, GLM-5.3 |
Coding-plan endpoint backing chat-smart / research-deep |
Subscription shims (OpenAI-compatible proxies over plans)
| Plan |
Exposed as |
Pattern |
| ChatGPT plan |
gpt-5.6-luna / sol / terra and friends |
Plan proxy holds plan credentials, refreshes tokens, serves /v1 to LiteLLM |
| Z.ai coding plan |
glm-5.2-subscription |
Plan endpoint with raw Bearer auth |
| Kimi CLI plan |
kimi-k2.5, kimi-cli, kimi-opencode-harness |
OAuth-refreshing shim |
| OpenCode Go |
*-opencode-go aliases |
OpenAI-compatible plan endpoint (24 models behind one key) |
Aggregators (pay-as-you-go)
| Aggregator |
Models via catalog |
Role |
| OpenRouter |
Claude Sonnet 5, Grok 4.3/4.5, GPT-5.5, MiniMax M3, GLM-5.2, DeepSeek V4 Flash, Qwen 3.6/3.7, Gemma 4, Mistral Nemo, Gemini 3.5 Flash, Kimi K2.6 |
PAYG mirrors + general long tail |
| Nous Research |
DeepSeek V4 Flash |
Alternative direct route |
| OpenAI (direct) |
Vision mirror, embeddings |
Separate credit pool for special-purpose models |
Capability providers (via the Provider Broker)
| Capability |
Providers |
Failover chain |
| Web search |
SearXNG (self-hosted), Serper, Brave, Tavily |
searxng → serper → brave → tavily |
| Research |
Perplexity |
sonar / sonar-pro only |
| Scraping |
Crawl4AI (self-hosted), Firecrawl |
crawl4ai → firecrawl |
| Speech-to-text |
Deepgram, OpenAI Whisper |
deepgram → openai |
| Text-to-speech |
ElevenLabs |
— |
| Voice calls |
Telnyx |
Signed webhooks, call control |
| Browser automation |
Browser Use (managed), self-hosted browser service |
Self-hosted by policy |
| Image generation |
Higgsfield |
Degraded-capability reporting |
| MCP services |
Scrapling and other registered servers |
Host allowlisted |
| Agent integrations |
Composio |
Least-privilege fixed operations |
Why so many providers?
- Resilience: a plan outage becomes a routing change, not an outage. The
platform regularly lives through provider incidents without users noticing.
- Cost: the same workload can be served at wildly different prices; tiers
and mirrors let policy pick the cheapest rung that satisfies the
requirement.
- Evaluation: running the same agent across vendors produces real
comparative data — which model is actually better for this workload is
decided by measurement, not marketing.
Sourcing discipline
- Each provider is pinned in the alias catalog with its exact endpoint, key
source, and documented quirks (e.g., endpoint differences between a plan
API and a general API; user-agent gates; output-limit mapping).
- Subscription-backed aliases are explicitly routed through the outer agent
chain's fallback order — never through invisible proxy-level fallbacks — so
the failover behavior is deterministic and reviewable.
- Keys live in Infisical and are projected into the proxy and broker
deployments only.