A local-first autonomous data analysis agent. Investigates, executes real code in secure sandboxes, self-corrects on errors, and independently verifies conclusions.
Manager plans & reasons: Formulates hypotheses and chooses the next analytical move.
Worker generates code: Writes targeted Python using DuckDB, Polars, or pandas.
AST guard screens code: Prevents dangerous system operations or interpreter traversal.
Execution in sandbox: Runs inside OS-contained subprocesses (Landlock/seccomp on Linux, sandbox-exec on macOS, Job Objects on Windows) or Docker.
Real feedback & self-correction: The manager inspects real output and tracebacks, self-correcting mistakes automatically.
Trust verification: Re-derives headline numbers by an alternative route and flags any ungrounded figures.
| Capability | Local mode | Cloud permitted |
|---|---|---|
| Privacy guarantee | 100% on-device — zero data egress | External LLM API calls permitted |
| Model backends | Ollama, LM Studio, local vLLM | OpenAI, Anthropic, custom gateways |
| Embeddings | Local sentence embeddings (embeddinggemma) | Remote embedding APIs |
| Network egress | Denied by default at the sandbox layer | Outbound package install on demand |