Self-hosted AI runtime

AI agents that work on your real machines.

PawFlow connects AI agents to your files, terminals, browsers, desktops, services, and durable workflows through controlled Relays.

Install in 5 minutes View on GitHub

Open source · Self-hosted · MIT licensed

CLIENTS

WEBCLIVS CODEAPI

PAWFLOWShared runtimeagents · context · flows · auth

MCPA2AAG-UI

CONTROLLED BOUNDARYRelayfilesystem · shell · browser · screen

YOUR INFRASTRUCTURE

REPODESKTOPGPUSERVICES

Your runtime coordinates. Relays execute where the work already lives.

01 / HOW IT WORKS

A controlled path to real infrastructure.

Clients share one runtime. Relays put execution next to the files, tools, and machines that matter.

USE IT FROM

WebCLIVS CodeAndroidTelegramAPI

CONTROL PLANE

PawFlow Runtime

agents · context · flows · auth · permissions

EXECUTION BOUNDARY

Controlled Relay

read-only or writable · container or host · explicit tools

WORK WHERE IT LIVES

MachineRepoDesktopBrowserServicesGPU

AI agents on your infrastructure, not in someone else’s sandbox.

02 / WHY PAWFLOW

One runtime. Three ideas.

The capabilities matter because they share the same architecture.

Work where your infrastructure already is

Connect agents to real machines through Relays instead of moving projects into disposable cloud sandboxes.

Explore Relays

One durable runtime

Models, agents, tools, memory, skills, conversations, and permissions persist across clients and sessions.

See the runtime

Reason with agents. Run with flows.

Use agents where reasoning helps. Move repeatable execution into flows with queues, retries, checkpoints, and backpressure.

Explore Flows

03 / SEE IT WORKING

Three proofs, not thirty promises.

RELAY / MCP

Ask ChatGPT about your real project.

ChatGPT → MCP → PawFlow → Relay → MyWorkspace.

No ZIP.
No stale copy.
Read-only when you want.

Connect a live workspace

A desktop session running through a PawFlow Relay

REAL MACHINE

Let an agent operate where the tools are.

Filesystem, terminal, browser, desktop, local services, and GPUs through one controlled boundary.

See Relay capabilities

An agent designs a deterministic PawFlow workflow

AGENT → FLOW

Design with an agent. Execute with a flow.

Use reasoning to build the work, then run it with explicit structure and crash recovery.

See durable Flows

04 / WORKS WITH YOUR STACK

Bring the model. Keep the runtime.

Mix providers and clients without rebuilding the system around each one.

MODELS

OpenAI · Claude · Codex · Gemini · DeepSeek · compatible APIs

PROTOCOLS

MCP · A2A · AG-UI

CLIENTS

Web · CLI · VS Code · Android · Telegram · API

INFRASTRUCTURE

Docker · VPS · workstation · remote machines · GPU hosts

OPEN SOURCESELF-HOSTEDMIT8,900+ TESTSMULTI-PROVIDERMULTI-MACHINE

Explore all features

05 / PRODUCT CATEGORY

Capabilities that usually live apart.

PawFlow combines a coding harness, workflow engine, real-machine execution, durable context, and interoperability because they share one runtime.

Capability	Coding harness	Cloud agent	Workflow engine	PawFlow
AI coding and tools	Yes	Yes	Limited	Yes
Real local machines	Usually	Usually no	Workers	Via Relay
Durable agent context	Usually	Yes	No	Yes
Deterministic workflows	No	Limited	Yes	Yes
Multi-model runtime	Varies	Varies	Limited	Yes
Multi-machine execution	Limited	Cloud VMs	Workers	Yes
Self-hosted control plane	Sometimes	Rarely	Often	Yes
MCP / A2A / AG-UI	Varies	Varies	Limited	Yes

06 / RUN IT YOURSELF

Run PawFlow on your infrastructure.

Install the runtime, connect a Relay, and give your first agent a real task in about five minutes.

Install PawFlow Read the docs View GitHub