The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
description: >- One switch, enforced in the Rust core. Local-only mode blocks cloud model calls, network tools, web search, integrations and cloud embeddings, and allows only on-device runtimes. Voice is the exception, because there is no local speech-to-text engine. icon: lock
Privacy mode
Most assistants' privacy story is a paragraph in a system prompt. OpenHuman's is an enforcement point in the Rust core.
The [privacy] config block defines three modes:
| Mode | What it means |
|---|---|
standard (default) |
Normal operation. Managed cloud routing, BYO providers and local models are all available. |
local_only |
No inference leaves the device. Every external chat provider is refused when it is built: the managed cloud, BYO cloud keys, even CLI delegates like Claude Code. Network tools, web search, integrations and cloud embeddings are refused at their egress points. Only local runtimes pass: Ollama, LM Studio, MLX and local OpenAI-compatible endpoints. Voice is the one exception, because no local speech-to-text engine exists. |
sensitive |
The base for an upcoming PII-aware tier (detection, redaction, destination disclosure). Today it behaves like standard. |
Why enforcement matters
Privacy mode is not a policy the model is asked to follow. The first check lives in the inference provider factory (crates/openhuman-core/src/inference/provider/factory/). Under local_only, the core refuses to build an external provider at all. The error names the blocked provider and tells you how to fix it: switch to a local model, or change the mode in Settings.
That makes the guarantee independent of prompts, agents, tools or upstream bugs. If any code path tries to reach a cloud model in local-only mode, it cannot get a client.
local_only is also enforced at the egress points where the agent sends your data anywhere else (crates/openhuman-core/src/security/egress/enforce.rs):
| Egress point | Behavior under local_only |
|---|---|
Network tools (http_request, web_fetch, curl) |
Refused, with a policy-blocked message in the tool result |
| Web search (Exa, Tavily) | Refused the same way |
| Composio tool calls and integration requests | Refused with an error |
| Cloud embeddings | Refused with an error |
| Local runtimes (Ollama, LM Studio, MLX, local OpenAI-compatible) | Always permitted, nothing leaves the device |
One deliberate exemption: backend control-plane round-trips keep flowing. These are sign-in, session, team, billing, and integration connection management and catalog reads. Blocking them breaks the app for no privacy gain. They carry auth tokens, ids and routing metadata, never user content.
Speech-to-text is not covered by local_only. There is no local speech-to-text engine, so dictation and transcription leave the device through the hosted OpenHuman STT proxy or the third-party STT provider you configured. Disable the voice component if your threat model requires all speech processing to stay local.
Privacy mode governs data egress. It is separate from the autonomy tiers (readonly, supervised, full), which govern what the agent may do. You can run a fully autonomous agent that never sends a byte of inference off-device.
Pairing it with local models
Local-only mode works with local AI. OpenHuman does not install runtimes or download models, so you set these up yourself before turning it on:
- Chat and reasoning on a runtime you run (Ollama, LM Studio, MLX, OMLX, or another OpenAI-compatible server), with the models you pulled, added as a provider under Connections → LLM.
- Piper text-to-speech, if you install the Piper binary and a voice yourself and set
PIPER_BIN. - Memory on a CortexDB endpoint you run yourself. The hosted TinyHumans engine is a cloud call, so local-only mode blocks it.
With local-only on, a workload routed to a runtime that is not running, or to a model you have not pulled, fails instead of falling back to the cloud.
See Use OpenHuman with a local model for a full local setup, and Keep sensitive data private for a broader privacy walkthrough.
See also
- Privacy and security: the full trust model (approval gate, sandboxing, path roots, command classification).
- OS keyring and secret storage: where credentials live.
- Local AI: the on-device model runtimes.