PawFlow

Product

One runtime from reasoning to reliable execution.

PawFlow is a self-hosted AI harness and runtime that connects agents to your real machines, shared context, and durable workflows.

Run PawFlow in 5 minutesSee the architecture
PawFlow runtime connecting clients to real infrastructure
Clients and external systems share one control plane; Relays execute next to the work.

Architecture

A durable control plane with a controlled execution boundary.

Models can change. Clients can change. The runtime keeps the agents, state, policies, and workflows together.

Clients
Web · CLI · VS Code · Android · Telegram · API
PawFlow Runtime
agents · context · flows · auth · permissions
Relay
filesystem · shell · browser · desktop · services
Your infrastructure
workstation · repo · VPS · GPU host

Shared durable runtime

The model is a component. The runtime is the product.

01

Persistent context

Memory, knowledge, diary, todos, scratchpads, project context, and conversations keep the right lifetime across sessions.

02

Reusable capability

Agents, tools, skills, providers, and permissions are configured once and shared across clients.

03

Continuous workspace

Continue the same conversation from Web, terminal, VS Code, Android, Telegram, API, or an embedded application.

Agents + flows

Use reasoning where it helps. Use structure where reliability matters.

Agents explore, inspect, code, delegate, and decide. Flows schedule, route, retry, transform, persist, recover, and deliver.

Agent designsExplore an unfamiliar task, use real tools, and shape the right procedure.
Flow executesMake the repeatable path visible through tasks, queues, checkpoints, retries, and backpressure.
Runtime observesInspect progress and recover execution without throwing away durable state.
Explore the Flow engine →
Agent work becoming a durable PawFlow Flow
Agent → Flow Editor → deployed Flow → Runtime Viewer.
External AI clients reaching a workspace through PawFlow MCP and a Relay
The same runtime can be exposed to other AI systems without duplicating its state.

Interoperability

One runtime, several interfaces.

MCPTools and context for LLM clients, including read-only live workspaces.
A2AAgent-to-agent discovery and interoperability.
AG-UIAgent-to-application and frontend interaction.
View integrations →

Own the whole path

Run agents on your infrastructure.

Install the self-hosted runtime, connect a Relay, and give an agent its first real task.

Start the quickstart