The University of AI Universitas AI
Socrates
Socrates Agent SDK

The model-agnostic alternative to vendor AI SDKs — a stateful Agent runtime. Import your SKILL.md and MCP tools. Switch models with one variable. Deploy anywhere, including air-gapped.

Contact Sales View Tiers Documentation
Provider Config
SKILL.md
Docker Deploy
Closed SDK · v0.1 · Model-Agnostic
# Point at any inference endpoint — swap in one line # OpenAI INFERENCE_PROVIDER_TYPE=openai OPENAI_API_KEY=sk-... DEFAULT_MODEL=gpt-4o-mini # Azure GovCloud (IL4/IL5) INFERENCE_PROVIDER_TYPE=azure-openai AZURE_OPENAI_ENDPOINT=https://<resource>.openai.azure.com/ DEFAULT_MODEL=gpt-4o # Ollama — fully air-gapped, no outbound INFERENCE_PROVIDER_TYPE=ollama OLLAMA_BASE_URL=http://localhost:11434/v1 DEFAULT_MODEL=llama3 # Session engine, tool graph, skills — identical in all three
# .skills/due-diligence/SKILL.md — portable workflow # Runs on any provider. $ARGUMENTS substituted at activation. --- name: due-diligence description: Run structured due diligence on a company or fund. version: 1.0.0 allowed-tools: web_search, web_scrape --- ## Step 1: Profile Retrieve a comprehensive profile of $ARGUMENTS — founding, leadership, funding history, core product, customer base. ## Step 2: Technology Assessment Assess the technical depth of $ARGUMENTS — core technology, defensibility, how it differs from competitors. ## Step 3: Risk Screen Screen $ARGUMENTS for litigation, regulatory actions, negative press, and key-person concentration risk. ## Step 4: Verdict Synthesize a verdict across all dimensions from prior steps. # Activate via REST — kernel auto-advances through steps PATCH /api/sessions/sess_abc123 { "metadata": { "skillMode": { "skillId": "due-diligence", "subject": "Stripe", "step": 1, "totalSteps": 4 } } }
# Build once — source never in the runtime image docker build -t socrates-kernel . # Run — OpenAI docker run -p 7001:7001 \ -e OPENAI_API_KEY=sk-... \ -v /your/skills:/app/.skills \ socrates-kernel # Run — fully air-gapped (Ollama on-prem) docker run -p 7001:7001 \ -e INFERENCE_PROVIDER_TYPE=ollama \ -e OLLAMA_BASE_URL=http://host.docker.internal:11434/v1 \ -e DEFAULT_MODEL=llama3 \ -v /your/skills:/app/.skills \ socrates-kernel # No source files in image. Skills mount from host. # Postgres session store included in compose.

What the Kernel Provides

Production-hardened infrastructure. The model is your decision.

Works With Any Inference Provider

OpenAI, Anthropic, Google, Groq, Azure OpenAI, Ollama, or any OpenAI-compatible endpoint. One environment variable selects the provider. The session engine, tool graph, and skill runner behave identically regardless of which model is underneath.

SKILL.md Runtime

Execute plain Markdown skill files — the same format as Claude Code skills. Multi-step workflows activate with a single API call; the kernel drives each step and advances automatically. Skills are yours. They travel to any model.

Stateful Session Engine

Postgres-backed conversation history, resumable across requests. Hybrid vector + keyword retrieval over session context. Sessions carry their own tool graph, persona config, and skill state — persistent and portable.

MCP Extensibility

Full MCP client over SSE, streamable HTTP, and stdio. Register any MCP server at runtime — the kernel discovers tools automatically and binds them into the session tool graph. Your domain layer stays outside the kernel.

Isolated Deployment

Multi-stage Docker build — compiled output only, no source in the runtime image. Mount your skills directory at runtime. Pair with on-prem Ollama for a fully offline stack. Designed for environments where data sovereignty, network isolation, or regulatory compliance is non-negotiable.

SSE Streaming

Token-level streaming with mid-stream tool invocation. Server-sent events over a single persistent connection. Built-in tool start/result/done event types. Resilient — guards against write-after-close across all tool paths.

Inference Providers

Agent behavior is identical regardless of which model is underneath.

Cloud
OpenAI
GPT-4o, GPT-4o Mini, o3
Cloud
Anthropic
Claude Sonnet, Opus, Haiku
Cloud
Google
Gemini 2.0, 2.5 Pro
Cloud · Fast Inference
Groq
Llama 3, Mixtral, Gemma
Secure Cloud · GovCloud
Azure OpenAI
IL4/IL5, FedRAMP High, private VNet
Secure Cloud · GovCloud
AWS Bedrock
GovCloud, Claude, Llama, Titan
On-Prem · Isolated
Ollama
Llama 3, Mistral, Qwen, any GGUF
Custom
Any OpenAI-Compatible
vLLM, Together, HF TGI, private endpoints

What the agent engine doesn't include — and why

The SDK is the runtime layer: session engine, tool graph, skill runner, streaming. Domain intelligence lives above it. Chain-of-Provenance, Constellation, Auto Analyst, and the AI Scientist expert network are Universitas AI platform features.

This is intentional. The agent engine runs anywhere. The platform is where primary sources, expert networks, and provenance chains live. Use both together for the full stack, or the SDK alone against your own data layer.

Contact us to discuss platform access alongside SDK licensing.

SDK Access

Closed distribution. Issued on request. Source is never included in the container SKU.

Secure Container
Regulated · Mission-Critical · On-Prem
Contact Sales
For organizations that require data sovereignty, network isolation, or bring-your-own-cloud
  • ✓ Compiled Docker image — no source
  • ✓ Runs fully offline with on-prem Ollama
  • ✓ Azure GovCloud (IL4/IL5, FedRAMP High)
  • ✓ Private VNet and air-gapped configs
  • ✓ Stateful session engine + Postgres
  • ✓ SKILL.md runner — bring your own skills
  • ✓ MCP client — wire your own tool servers
  • ✓ SSE streaming
  • ✓ Deployment support + SLA
Contact Sales
Source License
Platform Builders · Systems Integrators
Contact Sales
For organizations embedding the agent engine in their own product or platform
  • ✓ Full TypeScript source — all agent modules
  • ✓ All Secure Container capabilities
  • ✓ White-label rights under agreement
  • ✓ Open interface specs (REST, MCP, SKILL.md)
  • ✓ Integration engineering support
  • ✓ Platform API access available separately
  • — Examples and export modules: optional add-on
Contact Sales

All tiers include open interface specifications: REST API, SKILL.md format, MCP registration, and InferenceProvider config schema.