Architecture

High-level shape of the OpenHuman system (desktop shell, Rust core, Memory Tree, agent loop). Pointer to the deep developer architecture in the repo.

OpenHuman is open-sourced under GNU GPL3. This page is the high-level shape of the system; the deep developer architecture lives in deep architecture reference in the repo.

The shape

OpenHuman is a React + Tauri v2 desktop app with a Rust core that does the heavy lifting.

┌──────────────────────────────────────────────────┐
│ Tauri shell (app/src-tauri/) │
│ • windowing, OS integration, sidecar lifecycle │
│ • CEF child webviews for integration providers │
└──────────────────────────────────────────────────┘
 │ JSON-RPC (HTTP) ↕
┌──────────────────────────────────────────────────┐
│ Rust core (`openhuman` binary, `src/`) │
│ • Memory Tree pipeline │
│ • Integration adapters + auto-fetch scheduler │
│ • Provider router (model routing) │
│ • TokenJuice compression │
│ • Native tools (search, fetch, fs, git, …) │
│ • Voice (STT in, TTS out, Meet agent) │
└──────────────────────────────────────────────────┘
 │
┌──────────────────────────────────────────────────┐
│ React frontend (app/src/) │
│ • Screens, navigation │
│ • Talks to core over `coreRpcClient` │
│ • No business logic - presentation only │
└──────────────────────────────────────────────────┘

Where logic lives:

• Rust core. all business logic. Memory Tree, integrations, model routing, tools, voice. Authoritative.

• Tauri shell. windowing, process lifecycle, IPC. A delivery vehicle, not where features live.

• React frontend. UI and orchestration. Calls into core via JSON-RPC.

Data flow

1. Connect. OAuth into a integration. Backend stores the token; core never sees it in plaintext.

2. Auto-fetch. Every twenty minutes the scheduler walks every active connection and asks each native provider to sync.

3. Canonicalize. Provider output (an email page, a GitHub diff, a Slack channel dump) is normalized into provenance-tagged Markdown.

4. Chunk. Markdown is split into ≤3k-token deterministic chunks.

5. Store. Chunks land in SQLite (<workspace>/memory_tree/chunks.db) and as .md files in <workspace>/wiki/.

6. Score. Background workers run embeddings, entity extraction, hotness scoring.

7. Summarize. Source / topic / global summary trees are built and refreshed from the chunk pool.

8. Retrieve. When you ask a question, the agent queries the Memory Tree (search / drill down / topic / global / fetch).

9. Compress. Tool output and large source data go through TokenJuice before entering LLM context.

10. Route. The router picks the right provider+model for the task hint.

Privacy boundary

Stays on your machine:

• The Memory Tree SQLite DB.

• The Obsidian Markdown vault.

• Audio capture buffers and any local model state.

Goes through the OpenHuman backend (under one subscription):

• LLM calls (model providers).

• Web search proxy.

• Integration OAuth and tool proxying.

• TTS streaming.

See Privacy & Security for the full picture.

Open source

• Repo: github.com/tinyhumansai/openhuman. GNU GPL3.

• Issues and PRs are welcome. The project is in early beta.

• For contributors, the canonical developer guide is deep architecture reference.