The agentic ecosystem for open-weight models · launching Q4 2026

Run open-weight AI models.
Entirely on your machine.

OpenEng downloads, orchestrates, and runs open-weight (open-source) AI models on your own hardware — from any registry, in any major weight format, on the runtime that fits — isolated in a sandbox, never on your bare host. You drive it from the browser over an end-to-end-encrypted channel. Private and local, no telemetry.

Open the Web Console

Start your engine

$ openeng serve

Open-weight models On your machine No telemetry Encrypted channel

app.openeng.ai

Thin client · renders only · zero secrets

sealed

prompt in
tokens out

end-to-end sealed

Your machine

openeng serve · on your machine

Models

Secrets

Your code

100% on-device · never leaves the machine

OpenEng's broker only relays sealed frames it can't read. Models, secrets, and code never leave your machine.

OpenEng is an agentic ecosystem for open-weight model execution: it downloads, orchestrates, and runs open-weight (open-source) AI models entirely on your machine.

Built, tested, and real

1,092Agents

885Skills

754Templates

28On-device models

550Cells

44Loop strategies

2,433Engine tests passing

The architecture

One engine. Many surfaces. Sealed on the wire.

All intelligence and every secret live in a single local process. The browser, the CLI, and future IDEs are thin clients that only render — the only thing between them and your engine is OpenEng's broker, a blind relay that can't read a single frame.

The engine is local

A Rust engine that runs only on your machine and binds no inbound port — it dials out to OpenEng's broker. It owns the agent loop, models, memory, and every cloud call.

Clients hold nothing

The Web Console, CLI, and Desktop carry zero business logic and zero secrets. They capture input and render streamed event frames.

One sealed contract

A single versioned, encrypted contract drives every surface. The channel is sealed end-to-end — the broker relays it blind, never able to read a byte.

Why teams choose it

Privacy you can verify, not just trust.

The browser is physically incapable of leaking a secret — because it never receives one. The models run on your machine, and the wire has no field to carry your data off-box.

Browser

encrypted wire

Engine

Crosses the wire

Your prompt
Rendered output (tokens)

No field for it to travel through

Persona
Memory
Embeddings
Secrets
API keys
Your code

Encrypted end-to-end — the broker relays only ciphertexton your machine

The wire carries almost nothing
The channel carries only your request and the rendered tokens — no field for prompts, persona, memory, embeddings, secrets, or keys. There is nothing for the browser to leak.
Encrypted end-to-end
The browser-to-engine channel is encrypted end-to-end — the broker that relays it only ever sees ciphertext, even on your own machine.
Thin client, by construction
The browser bundle holds no model and no secret. It cannot leak what it was never built to hold.

The engine

Any registry. Any format. The runtime that fits.

A model is three independent choices: where the weights come from, what format they are in, and which runtime executes them — the format determines which runtimes can run it. Combine several models in one config — a role group — so each job runs on a right-sized model. Local models run isolated in a sandbox; hosted models route through one gateway.

Three axes, one matrix

Pull weights from Hugging Face, the Ollama library, ModelScope, or a direct URL — GGUF, Safetensors, GPTQ/AWQ, EXL2/EXL3, MLX, ONNX, or TensorRT — and run them on a compatible runtime, from llama.cpp and ONNX Runtime to vLLM, MLX, and Ollama.

28 on-device models

Open-weight models from ~1.3 GB to a 70B, auto-matched to your hardware and run isolated in a sandbox on your machine. Zero API keys.

Any registry, any format
Hardware-aware selection

Private inference, offline

With models on disk the engine answers, plans, codes, and reviews with no network call. Optional Metal / CUDA / Vulkan acceleration with a single build flag.

How the engine works

The product

The console runs in your browser. The work runs in a sandbox on your machine.

A thin React client that streams your own engine's work over a sealed channel — engine-owned threads, streaming replies, in-browser permission prompts. No native install for the interface.

app.openeng.ai

Migrate the users table to Postgres and update the data layer.

Planning · 4 steps

Map the current schema
Write the migration
Update the data-access layer
Update tests + docs

Ask anything…complexity 84% · planned

Zero-click connect

Sign in once, run openeng serve, and the console connects the instant your engine comes online — through OpenEng's broker, nothing to paste.

No modes — just type

The engine scores each request's complexity and self-routes plan-vs-execute. You never pick 'chat vs goal vs coding'.

Threads, everywhere

Engine-owned threads are shared across the Web Console, CLI, and Desktop — the same account, the same history.

The control plane

Author once. Publish a version. Run anywhere you sign in.

Configure your whole agent platform in the browser, freeze it into an immutable id@version, and every machine you sign in to pulls the exact same setup. No drift.

16 config domains

Models, agents, execution flow, schedulers, templates, cells, plugins, guardrails, MCP servers — Build, Connect, Govern, Deploy, one console.

Immutable, versioned

Publish-once with strictly-higher SemVer. See exactly what your engine will run before you ship it — mirror, not source.

Secrets never touch it

The control plane stores ${ENV} references only — your actual secrets resolve from the local environment, never the cloud.

Explore the Console

The ecosystem

Don't start from a blank prompt. Start from 3,300+ building blocks.

Browse and snap together agents, skills, templates, models, plugins, cells, and loops — then publish a config your engine just runs. Compose, don't code.

Agents1,092

A 3-level org chart — 282 orchestrators delegating to 810 single-purpose agents.

agents.openeng.ai

Skills885

Named, invocable playbooks the engine runs on demand, across 18 categories.

skills.openeng.ai

Templates754

Prompt blueprints prepended above your persona, across 28 segments.

templates.openeng.ai

Cells550

AWS / Azure / GCP connectors — read-only by default, inside your own IAM.

cells.openeng.ai