Ask your data,watch it reason
An autonomous data analysis agent powered by a Tri-Model architecture (Manager, Worker, Embeddings). Runs sandboxed Python, self-corrects tracebacks, and verifies its own findings — locally on your machine.
brew install Wizard-AIA/wizard/wizardAn agent that
investigates.
Real analytical questions are rarely one step — a dirty join key or a column that means three things only shows up once you look.
Tri-Model Architecture
Dedicated roles for Manager (DAG orchestration & reasoning), Worker (executable Python authoring), and Embeddings (vector context retrieval). Mix and match Local (Ollama), Hybrid, or Cloud models per role.
Upload.Ask.Verify.
It runs on
your machine.
Generated code runs in a subprocess by default — its own memory ceiling, a per-step timeout, a working Stop button. Docker is opt-in, not a fallback.
host — a sandboxed subprocess, the default. docker — a container per session, opt-in. inprocess — used only when spawning is forbidden.
Numbers you
can check.
Bring your
own model.
Nothing is hardcoded — the app uses whatever your provider actually has, and you can assign a different model, on a different backend, to each role.
Autonomous,
not uncontrolled.
Generated code is untrusted by default. Four layers apply before it ever touches your filesystem or the network.
Static code guard
An AST policy check rejects restricted imports and dynamic execution before anything runs.
OS-native containment
Landlock + seccomp, sandbox-exec, or a Windows Job Object — writes confined, network denied.
Container isolation
Opt into Docker for cap_drop=ALL, no-new-privileges and per-execution timeouts.
Scoped filesystem
Each session reads and writes only its own workspace directory.
A REST API.
And a WebSocket.
The same protocol the web UI uses to stream reasoning, code, stdout and the answer is open for anything else you want to build against it.
Typed event protocol
session · reasoning_delta · code · stdout · final — every WS & SSE frame is a fixed typed contract.
Streams as it thinks
Reasoning, code, stdout and the verified answer arrive as separate delta streams, token by token.
OpenTelemetry & Prometheus
Built-in distributed tracing, TTFT latency tracking, and live /metrics endpoint ready for enterprise observability.
OpenAPI, generated
Interactive Swagger docs at /docs; TypeScript contracts generated straight from the schema with zero drift.
In the project's own words.
Two small Ollama models are enough to be useful. Nothing is sent anywhere.
Wizard
Local-first, README.md
Pick your
setup.
Local-only
Zero outbound bytes. All reasoning, code generation, and embeddings stay on your machine.
- Ollama or LM Studio for Manager, Worker & Embeddings
- Local vector embeddings (nomic-embed-text / all-minilm)
- Zero bytes egress, zero API keys required
Hybrid
Keep local models for privacy & speed, escalate to cloud reasoning or embeddings when needed.
- Local models for execution with cloud escalation
- One cloud API key (Gemini, Claude, OpenAI)
- Automatic schema redaction & granular role assignment
Cloud-only
Zero weight downloads needed — state-of-the-art reasoning, coding, and native batch embeddings.
- Gemini, Anthropic, OpenAI, or custom gateway
- Native Google batchEmbedContents & cloud embeddings
- Auto-discovery of active models with zero configuration
Ready to ask
your data something?
Local-first, no API key required. Two small Ollama models are enough to get started.
BSD 3-Clause · runs entirely on your machine
