Devices are not servers.
App termination, memory pressure and limited background execution interrupt agent loops. Completed work should not disappear with the process.
Open-source infrastructure for edge AI
Build multi-step agents that preserve their progress, recover from interruptions, and coordinate execution across mobile devices, edge hardware and the cloud.
In development · Open‑source runtime planned · Model‑agnostic · Source on GitHub
The agent has analyzed the image and recorded its first tool result.
Interactive concept simulation · not a live runtime
01 / The missing layer
Local models make edge intelligence possible. Multi-step agents introduce a different challenge: preserving decisions, tool results and execution progress when devices disconnect, suspend or terminate.
App termination, memory pressure and limited background execution interrupt agent loops. Completed work should not disappear with the process.
Local reasoning and tools should keep working without a reachable orchestration service, when the device permits execution.
Handing work to remote models and tools requires persistent dispatch, retry safety and a consistent execution history.
02 / Developer experience
A high-level agent API for fast adoption. A durable workflow API for complete control.
API Preview · TypeScript · initial target: React Native.
import { createAgent } from "@larkx/agent";
import { cactusAdapter } from "@larkx/cactus";
const agent = createAgent({
name: "inspection-agent",
model: cactusAdapter(model),
tools: [vision, records, report],
durability: true,
});
await agent.run({
input: "Inspect this equipment"
});
// On next app launch
await engine.recover();03 / Technology stack
Models provide intelligence. Agent frameworks decide what to do. LarkX is designed to make those multi-step decisions and actions durable across devices and the cloud.
* Product Preview. Illustrative stack, not strict dependencies: some platforms span multiple layers. Named third-party technologies are examples, not confirmed integrations or partnerships.
MODEL-AGNOSTIC · LOCAL-FIRST04 / Architecture
Unlike server-coordinated workflow engines, LarkX is designed to embed orchestration and execution history directly on the device.
Run locally. Manage globally.
The runtime owns the workflow journal. Models and tools remain replaceable. Remote steps are coordinated through a managed control plane, using the customer's own compute.
Recovery happens when the OS allows execution. Cloud handoff requires connectivity, compatible handlers and permission to transfer data.
EDGE / EMBEDDED
CLOUD / CUSTOMER
← ONE DURABLE EXECUTION ID →
Product Preview · LarkX Runtime
Low-level workflow primitives: named steps, checkpoints, replay, and a device-local execution journal.
05 / Competitive landscape
Temporal made backend workflows durable. LarkX is bringing durable execution to the devices where agents increasingly run.
| Existing solution | Primary focus | LarkX's proposed distinction |
|---|---|---|
| Temporal (opens in a new tab) | Server-coordinated durable workflows | Embedded offline agent orchestration |
| DBOS (opens in a new tab) | Database-backed backend execution | Device-local SQLite and mobile lifecycle |
| Golem (opens in a new tab) | Durable WebAssembly agent runtime | Embedded mobile lifecycle and offline-first coordination |
| Restate (opens in a new tab) | Durable service and agent execution | Device-local orchestration without a reachable service |
| LangGraph (opens in a new tab) | Agent graphs and checkpointing | Device-first runtime and hybrid handoff |
| RunAnywhere (opens in a new tab) | Local inference, agent loops and fleet tooling | Focus on durable multi-step execution and recovery |
| PowerSync (opens in a new tab) | Offline database synchronization | Execution history and operation recovery |
Comparison of architectural focus, not an exhaustive feature audit. Existing systems can be integrated with LarkX. On small screens the table scrolls sideways.
Model-agnostic by design
Bring the inference engine you already use. LarkX is designed to preserve the decisions, tool results and execution state around it.
Illustrative integration targets, not shipped integrations or partnerships. Ollama is primarily a desktop/server runtime; availability varies by device.
06 / Business model
Free embedded execution drives developer adoption. A managed hybrid orchestration platform creates recurring revenue without reselling GPU compute.
Planned open-source core
Free local execution. No mandatory cloud dependency.
Planned commercial platform
Managed orchestration for distributed agent execution.
07 / Initial markets
Starting with mobile and desktop developers, expanding to enterprise, robotics, and IoT fleets where reliability and offline access matter.
Analyze equipment locally, gather evidence in dead zones, delegate complex reasoning when connectivity returns.
Multi-step voice, vision and planning agents that recover after app closure rather than losing progress.
Coordinate high-level AI workflows across constrained hardware and remote workers, separate from real-time safety control.
08 / Roadmap
SQLite replay, React Native, process-death recovery and Cactus agent demo.
Persistent dispatch, cloud workers, reconciliation and hosted execution history.
Cross-device coordination, capability-aware scheduling and fleet management.
Building the next execution layer
We're looking for developers building on-device agents, design partners with real hybrid execution challenges, and investors who share our vision.