Local-first · Apple Silicon · BYO key
The context layer your Mac was missing.
Human turns what you see and do into structured, local memory and puts it behind one hotkey — so any AI acts with your full context instead of starting from zero. Raw data never leaves your machine.
macOS 26+ · on-device models · export everything as SQLite + Markdown + JSON
Frontier models are converging on price and capability. The bottleneck isn't intelligence anymore — it's context. Human is the local-first layer that captures, structures, and serves the context that never makes it into a prompt: your screen, your last hour, the project, the decisions. On your machine, on your terms.
Architecture
The whole pipeline runs on your machine.
Capture → memory → retrieval → answer, with approval-gated actions on top. Here's each stage, with the real components named.
- 01
Perceive
Reads the text and structure of your active window (accessibility-first), with scoped screenshot + on-device OCR only where there's no readable text. A perception gate drops blacklisted apps and secure-input fields before anything hits disk.
- 02
Distill
Raw activity is distilled into structured cards — decisions, people, projects — with entities resolved and duplicates merged. On-device Foundation Models or a redacted Claude pass; every output is grounded against its source before it's kept.
- 03
Store
Everything lands in a local, encrypted SQLite database with full-text and vector indexes. No cloud store, no account required. Per-kind retention; a janitor thins old frames.
- 04
Retrieve
A query router normalizes time expressions and runs hybrid retrieval — FTS + vector candidates fused with reciprocal-rank fusion — then a referent resolver figures out what “this” and “that” point to.
- 05
Reason
The assembled context passes the redaction gate before TLS, then the model. Answers stream back with provenance linking to the exact moments and cards they came from.
- 06
Act
Actions use a typed Mac vocabulary through a deterministic allow / ask / deny classifier and a manual approval broker. AX-first targeting with OCR/grid fallbacks. Nothing side-effecting runs unapproved.
Local-first
Private by architecture. We hold nothing.
- Perception gate drops blacklisted apps and secure-input before disk — not encrypted later, never written.
- Redaction gate masks cards, IBANs, keys and tokens before any cloud call; per-app “never send” lists; every call audited.
- Encrypted at rest — AES-256-GCM with Secure Enclave-wrapped keys; store in a TCC-protected container.
- Own your data — export the full SQLite + Markdown + JSON vault; nuke crypto-shreds keys first.
Screen & activity
Text-first capture, selective frames
Local memory & retrieval
Encrypted · searchable · yours
Raw captures never cross this line. Cards, IBANs, secrets and blocked apps are masked before anything leaves.
Minimum context for an answer
Redacted text only — audited every time
Human API · MCP
Your context, available to every AI you use.
Human exposes your context through grant-scoped MCP tools — so Claude Desktop, Cursor, or your own agent can read exactly the slice you allow.
- Bound to 127.0.0.1 — unreachable from the network.
- Per-client keys in the Keychain; scoped, revocable, and audited.
- Redaction on every output; a Shared Context Folder for clients that can't reach localhost.
A portable, user-owned context layer future AI has to ask permission to use — the first piece of the Human API.
On your terms
The spec.
- Platform
- macOS 26+, Apple Silicon · notarized, outside the App Store
- Models
- Claude (Haiku) — bring your own key, or hosted access
- On-device
- Foundation Models, SpeechAnalyzer, Vision OCR, NL embeddings
- Data
- local encrypted SQLite · export as SQLite + Markdown + JSON
- Actions
- approval-gated · send/buy/pay/delete/publish blocked in v1
- Control
- pause, private mode, per-app cloud blocks, delete by range, nuke
Point your AI at your actual context.
Join the private beta. Apple Silicon, macOS 26+, bring your own key.