Getting Started

Stimulir Documentation

Stimulir is an adaptive AI stack for evidence-backed improvement. The console gives every team three workspaces, Engineering, Lab, and Compute, backed by the HybrIE runtime.

Latest release

v0.1.271 fixes managed vendor replay routing while retaining guided, signed Improve previews, tenant-authorized execution, and source-backed demo projects.

Create evaluation-ready projects from seven pinned open-source datasets with visible provenance, licences, and 150 real records per pack. Overlord controls can independently govern adapters, projectors, and adapter evaluation, while admin toggles update without reloading the page. Improve still builds privacy-gated immutable cohorts from tagged production traces. Start with Improve or the Coding Agent Skills guide.

Two API surfaces

Stimulir exposes two complementary APIs:

  • Console platform API, https://api.stimulir.com. Manages workspaces, API keys, BYOK credentials, usage, and billing, and serves OpenAI-compatible inference at /api/v1/inference/chat/completions. Platform endpoints authenticate with your session token; inference authenticates with hyb_* API keys. See Authentication & Workspaces.
  • HybrIE runtime API, the engine the Lab and Compute workspaces control. It serves an OpenAI-compatible HTTP API on port 8080 (gRPC on 9090) with local inference (Qwen3 / Qwen3-Coder via Candle on Metal or CUDA), training, evaluation, adapters, and compute orchestration. In BYOC deployments you run this runtime on your own nodes.

Latest capabilities

Improve

Turn captured production traces into a durable, budget-bounded prompt hill climb with evidence, lineage, and no automatic promotion.

Coding Agent Skills

Give Codex or Claude Code short operator prompts for connection, capture, evaluation, Improve, and human-gated promotion.

Stimulir Fusion

Call stimulir/fusion or stimulir/fusion-max to fan a prompt out to a model panel and return one judge-written answer.

Prompt management

Create versions, read by key and label, update metadata, archive versions, and preserve lineage for evals.

Data asset hub

Upload, ingest from traces, update metadata, bulk-stage, unstage, snapshot, and hand off curated assets to Lab.

Lab evaluation runs

Run prompt, data, endpoint, adapter, RL, and NIAH evaluations with durable run records and reports.

Python SDK

Add the SDK with uv and use StimulirClient for prompts, data assets, eval runs, and capabilities.

CLI paths

Run prompt, data, inference, training, adapter, eval, and compute workflows from the terminal.

Explore the docs

Engineering Workspace

OpenAI-compatible inference, API keys, BYOK credentials, usage metering, prompt versions, and data assets.

Lab Workspace

PEFT LoRA training, Doc-to-LoRA context internalization, durable eval runs, and hot-swap adapter serving.

Compute Workspace

GPU offers and instances, worker registration, and edge deployment across local, hybrid, and P2P modes.

CLI

Install the terminal command with uv tool, then run prompt, data, eval, inference, and compute workflows.

Next steps

  • Follow the Quickstart to install the CLI, create an API key, and make your first inference call.
  • Use the Python SDK or prompts and data assets commands to seed client prompts, ingest traces, stage datasets, and create Lab-ready snapshots.
  • Run Lab evals against prompt versions, staged data, inference endpoints, adapters, and RL policies before promotion.
  • Use Improve to diagnose captured production behavior and hill-climb a prompt; review the result before any separate promotion action.