Skip to main content

Start Here

AtomGradient Edge is a local-first developer platform for building private AI agents that run, learn, and coordinate on user-owned devices.

The Developer Preview ships on Apple platforms first. Android, Linux, HarmonyOS, Windows, robots, vehicles, and industrial devices share the same long-term technical core: local models, local learning artifacts, app-owned tools, and explicit compatibility gates.

Developer Preview

All Edge products are in Developer Preview. APIs may change between releases. Some repositories and Swift package dependencies may still require AtomGradient preview access. Pin package versions and validate on real devices after each upgrade.

Getting started

GoalGuideExpected result
Download, chat, then run a learning demoCLI learning demoLocal chat works, then a synthetic correction generates a Neural Imprint artifact with a hash-only comparison receipt.
Install the preview packageInstall Edge StudioThe edge CLI is installed from the edge-studio Python package.
Launch the local workbenchLaunch the Web UIedge studio runs Edge Studio at http://127.0.0.1:18842.
Build an iOS shellMinimal iOS appEdge Scaffold compiles as the smallest iOS reference app. Preview access required.
Integrate the Swift SDKSwift SDK setupEdge Kit loads a local model in an Apple-platform app.

First commands

Download a model and chat locally:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install edge-studio
edge models fetch qwen3.5-9b-4bit --source auto
edge demo chat --model qwen3.5-9b-4bit --interactive

After the base chat works, continue to the CLI learning demo to inspect a synthetic correction sample, generate a local Neural Imprint artifact, and compare before/after answer hashes.

Product stack

ProductWhat developers use it for
Edge StudioLocal workbench and CLI for model readiness, model fetch receipts, local learning demos, Neural Imprint generation, device management, benchmark, and export.
Edge KitSwift SDK for LLM, VLM, speech, model management, EdgeData, EdgeMesh, EdgeDataMeshBridge, EdgeSession, and EdgeUI.
Edge EngineNative on-device inference runtime. Packaged under Edge Kit; most apps do not import it directly.
Edge HaloPersonalization lifecycle layer: profile jobs, Neural Imprint capsule validation, restore orchestration, and compatibility gates.
Edge ScaffoldReference app and export template showing the recommended iOS integration pattern.

Privacy model

Edge is designed around user-owned compute:

  • Inference runs locally.
  • Training inputs, corrections, and conversation history remain app-managed local data.
  • EdgeMesh transfer is local-network and trust-gated.
  • Neural Imprint artifacts are compatibility-checked before restore and can be removed by the app.

Do not upload user transcripts, corrections, or profile artifacts to analytics, crash logs, or remote support systems.

Core concepts

ConceptDeveloper-facing meaning
Local-first inferenceModels, prompts, user data, and personalization artifacts stay on user-owned devices unless the user explicitly enables local mesh transfer.
Neural ImprintA local personalization artifact that lets a compatible base model restore a user-specific state without changing model weights.
App-owned toolsApps define their own tool schemas and action surfaces. Edge infrastructure should not embed app business rules.
EdgeMeshLocal-network trust, discovery, and device-to-device transfer for user-owned devices.
Fail-closed compatibilityPersonalization and model artifacts must match model identity, tokenizer/template identity, runtime version, and tool schema before restore.