Status: built and planned

CriogenAI is at 0.1: early, tested, and changing. The API may still change before 1.0. This page is kept accurate: if something is not marked built, do not rely on it.

By feature and language#

FeaturePythonTypeScriptRust
Agents with tools, limits, structured outputbuiltbuiltbuilt
Human approval (approve, reject, edit, respond)builtbuiltbuilt
Threads: saved conversations, retry, history, rewindbuiltbuiltbuilt
Streamingbuiltbuiltbuilt
Cancellation (asyncio / AbortSignal / token)builtbuiltbuilt
Storage: SQLite, Postgres, in memorybuiltbuiltbuilt
Encryption at rest, audit trail, retentionbuiltbuiltbuilt
Long-term memory (keyword search)builtbuiltbuilt
Long-term memory by meaning (embeddings)builtbuiltbuilt
Collections: documents searched by meaning, exact or HNSW index, as an agent toolbuilt (0.1.5)built (0.1.5)built
Workflows (graphs of steps)builtbuiltbuilt
Durable workflows that pause for a personbuiltbuiltbuilt
Hooks around model and tool callsbuiltbuiltbuilt
Agents as tools of other agents, agents as workflow stepsbuiltbuiltbuilt
Fallback models, retry policiesbuiltbuiltbuilt
Local trace viewer (every run on a page, on your computer)built (0.1.5)built (0.1.5; Node only)built
Evals: datasets, scorers, an LLM judge, comparisons, a CI gatebuilt (0.1.5)built (0.1.5)built
Tracing to Langfuse / OpenTelemetrybuiltbuiltbuilt
Tool input from Zod / Standard Schema–built–
Azure OpenAI and Foundry (API key, service principal, managed identity)built (0.1.6; not yet checked live)built (0.1.6; not yet checked live)built
Amazon Bedrock with API keys (Claude over Messages, others over Chat Completions)built (0.1.6; not yet checked live)built (0.1.6; not yet checked live)built
Google Vertex AI (Gemini, Claude; gcloud credentials, metadata server, tokens)built (0.1.6; not yet checked live)built (0.1.6; not yet checked live)built
OpenAI over the Responses API (reasoning with tools, nothing stored at OpenAI)built (0.1.6)built (0.1.6)built
Calling a model directly (Gateway: chat, stream, embed)builtbuiltbuilt
Patterns: sequential, parallel, loop, review and critique, coordinator, hierarchy, swarm, human reviewbuiltbuiltbuilt
Specialist agents you can extend and registerbuiltbuiltbuilt
Tools from MCP servers (stdio, Streamable HTTP; MCP 2026-07-28 and older)built (0.1.5)built (0.1.5; criogenai/web: HTTP only)built
Web search (Tavily, Brave) and page readingbuiltbuiltbuilt
In the browser and Web Workers: the engine as WebAssembly (criogenai/web)–built (0.1.4)–
Models written in your own code (FunctionModel; in Rust, the Model trait)–built (0.1.4; on Linux and macOS with the next engine release)built

About the new name and platforms: from 0.1.7 the packages are called criogenai (until 0.1.6: criogen). 0.1.5 to 0.1.7 were built on Windows, so Windows and the browser engine (criogenai/web) have everything above. The Linux and macOS engines for criogenai come with the next build, planned for 16 November; until then those systems use the old criogen package with the 0.1.3 engine (0.1.2 for Python on Intel Macs), where the features marked 0.1.4 or later are missing and openai: still uses Chat Completions.

The gateway#

FeatureStatus
OpenAI-compatible /v1/chat/completions (streaming too), /v1/modelsbuilt
Any upstream: OpenAI, Anthropic, Gemini, OpenRouter, Ollama, vLLMbuilt
Routes with fallback models, virtual keys, usage recordsbuilt
Anthropic-format /v1/messages (streaming too)built
Rate limits and token budgets per key (in memory)built
Pass-through mode, shared counters, budgets in moneyplanned

Platforms#

Tested on Windows x64 and Linux x64. Builds for Linux ARM64 and macOS (Apple Silicon and Intel) are set up but not yet tested. In the browser: tested in Chromium, Firefox and WebKit (Safari's engine), and in Web Workers. Edge runtimes (Cloudflare Workers, Vercel Edge) are not yet tested.