Open source · v0.0.1 & v0.0.2 available

Universal intelligence.
Trustworthy execution.
Human empowerment.

Origin Agent is a native, model-agnostic intelligent agent environment that works offline, keeps your data on your device, and treats trust as an execution property — not a slogan.

Powered by Origin AI & the Origin Universal Harness

  • Android 8+
  • 4 GB target
  • No telemetry
  • MIT license

Trust, engineered in

It does the work
behind the scenes.

Every consequential action moves through one explicit pipeline. You see what was planned, what you approved, what actually ran, and how the result was verified — with a signed chain of provenance you can check offline.

  1. 01IntentWhat you asked for
  2. 02PlanProposed steps
  3. 03AuthorizationYour explicit consent
  4. 04ExecuteBounded tool run
  5. 05VerifyOutcome checked
  6. 06PersistLocal, append-only
  7. 07ProvenanceHash-chained ledger
  8. 08ResultExplained to you

Built to adapt

Works with any model,
any runtime, any device.

Principles stay stable. Technology choices stay replaceable. No single provider, chip or model gets a permanent seat at the table.

Any model

Local deterministic intelligence is always available. Bring your own model — local or remote — and decide which one answers. No bundled weights, no forced favorite.

Explore the model registry

Any runtime

Optional adapters for llama.cpp, bitnet.cpp and OpenAI-compatible endpoints, with honest availability checks and clear attribution of every response.

See the architecture

Any device

Android 8+ and a 4 GB reference target (Galaxy S9), optimized for ARM64 and x86-64 — graceful on constrained hardware, at home on newer devices.

Read the requirements

Conversation is the interface

Starts with what you ask.

No setup wizard. No model download. Open Origin and talk. The deterministic local engine answers useful questions from the first launch, while suggestions help you discover what the app can do.

  • General-purpose chat, immediately usable
  • Honest engine status — you always know what is answering
  • Live trust badges on every response

Tools that stay in your control

Productive when you need it.

A conversation can become a workflow. Calculator, date and time, notes, and knowledge search are deterministic tools that work with zero setup and zero network. Capabilities can be enabled, inspected and consented to individually.

  • Capability catalog with clear descriptions
  • Per-tool consent controls
  • Local knowledge search across your saved material

Consent is not a checkbox

Asks before it acts.

Origin Agent separates a proposed action from an authorized one, and attempted execution from verified success. Consequential steps pause for your approval, then land in a tamper-evident trust ledger with a hash-chained provenance record you can verify offline.

Read the Constitution

Privacy you can inspect

Sovereign by default.
Never sees your passwords.

Memory lives on your device. There is no telemetry, no mandatory account, no required cloud and no subscription. No secret credentials are collected by the app — if you point it at a remote endpoint, you configure it yourself and stay in control.

  • Local append-only memory with user-controlled export
  • No ads, analytics or proprietary SDKs
  • Minimal permissions, honest capability descriptions

Memory that belongs to you

Learns how you like things done.

Searchable notes and facts accumulate locally, with explicit provenance for where they came from. One tap exports everything as portable JSON. Clearing is deliberate and confirmed — nothing disappears silently.

  • Structured, portable, interoperable knowledge
  • Source attribution on remembered facts
  • Clear-with-confirmation, always

Available in v0.0.2

Choose the model that answers.

The universal model picker lets you select and switch actual models — a local deterministic engine, optional bitnet.cpp or llama.cpp runtimes, or a web adapter. Add adapter offers built-in templates (OpenAI, OpenRouter, Groq, Together, DeepSeek, Ollama, LM Studio) as well as any OpenAI-compatible endpoint you trust. Web access is possible and enabled by default with a Settings switch, and no request happens until you select a web adapter. Conversations continue across switches, and every reply carries clear model attribution.

Strategic candidates, not a ranking

Origin AI evaluates models on evidence in the intended environment. Nothing earns permanent preference, and no weights are bundled with the APK.

  • BitNet b1.58 2B4T
  • LFM2.5-class
  • Optimized sub-1B
  • Qwen-class compact
  • Emerging quantization
View selection criteria

No gates, no gatekeepers

Open source.
Self-hosted. Sovereign.

Build it yourself, inspect every line, and install from a source you trust. Origin Agent ships under MIT with no Play Store dependency and no proprietary SDKs.

MIT licensed

Read, fork, audit and improve. The license is short and the source is public.

View license

Reproducible builds

Build the release APK from source with documented SDK, NDK and Gradle steps.

Build instructions

Not on Google Play

APKs come from GitHub Releases with published SHA-256 checksums. F-Droid metadata is included.

Get releases

Governed in the open

The Constitution defines the principles. The roadmap and evidence log decide what changes next.

Read the Constitution

Good questions

Frequently asked questions.

Still curious? Open an issue or read the docs in the repository — everything is public.

Ask on GitHub
What is Origin Agent?

A native, mobile-first, model-agnostic intelligent agent environment for Android. It combines conversation, local tools, persistent memory and a verifiable trust pipeline in one app. Origin AI is its intelligence foundation; Origin Agent is the product.

Can I use it today?

Yes. v0.0.2 (universal model picker, adapters and web access) and v0.0.1 are published on GitHub Releases with the Origin chat, Capabilities, Memory, Trust and Settings screens. Download the APK from a source you trust and verify the SHA-256 in the release notes.

Does it work offline?

Yes, fully. The deterministic local engine, calculator, date/time, notes, knowledge search and trust ledger all work with no connection. Optional model runtimes are user-supplied, and remote endpoints are opt-in.

Does it make network requests?

Only when you choose to. Web access is possible and enabled by default, but you can switch it off in Settings at any time. There is no telemetry, analytics or required cloud service: the app reaches the network only if you select a web adapter (a built-in template or your own OpenAI-compatible endpoint). Local models stay fully offline, and every reply tells you which model answered.

Where does my data live?

On your device, in local append-only memory. There is no mandatory account and no synchronisation. You can search it, export it as portable JSON, or clear it with confirmation.

Which models are supported?

v0.0.1 uses a deterministic local engine. v0.0.2 adds a picker for local deterministic, optional bitnet.cpp and llama.cpp models, plus web adapters — built-in templates (OpenAI, OpenRouter, Groq, Together, DeepSeek, Ollama, LM Studio) and any OpenAI-compatible endpoint you add yourself. Strategic candidates — BitNet b1.58 2B4T, LFM2.5-class, optimized sub-1B and Qwen-class compact — are evaluated on evidence, never assumed. See the model registry.

Is it on Google Play?

No, not by design for now. APKs are distributed through GitHub Releases, and F-Droid metadata is included in the repository. No Play Store dependency, no ads, no proprietary SDKs.

Origin Agent

Intelligence that works for you, answers to you, and stays with you. Open the repo, read the principles, or install the latest APK.

v0.0.2 available now · v0.0.1 also available