Open source. Self-hosted. Built for real work.

Self-hosted agent runtime for real infrastructure.

Run durable AI agents against your own files, tools, browsers, desktops, services, and workflows without moving the runtime into a vendor-controlled agent cloud.

Install PawFlow See how it works

Relay-backed tools Shared context Multi-provider agents Deterministic flows

Web UIchat, context, memory, terminals PawCodeterminal client, shared conversations Desktop Relayremote or local screen access Flowsscheduled DAGs with backpressure

Why PawFlow

Agents should work where your infrastructure already is.

PawFlow keeps the orchestration layer under your control. The server coordinates agents and conversations; relays execute tools next to the workspace; flows run repeatable work with explicit structure.

Your files stay reachable through relays.Agents can read, edit, search, run commands, use browser/desktop actions, and generate media through controlled tool routes.

Context survives restarts.Conversations, memory, knowledge graphs, diaries, project graphs, and FileStore outputs remain available across clients.

Providers are interchangeable.Use Codex app-server, Claude Code interactive, Antigravity/Agy, Gemini CLI, Anthropic, OpenAI, or compatible endpoints per agent or conversation.

PawFlow web chat and flow console preview — One runtime shared by web chat, PawCode, VS Code, API clients, and channel flows.

Architecture

A clear boundary between server, relay, agent, and flow.

PawFlow does not need permanent direct access to every machine. It routes work through connected relays and explicit services.

Clients
Web, CLI, VS Code, API

PawFlow Server
agents, context, flows, auth

Relay
filesystem, shell, screen, browser

Workspace
code, desktop, local services

Design with agents. Execute with structure.

Use agents for exploration, coding, decisions, and maintenance. When work becomes repeatable, turn it into a flow: CRON triggers, task DAGs, backpressure, checkpoints, retries, and explicit LLM calls only where modeled.

View the daily digest pattern

Webchat workspace

The conversation can open the same surfaces the agent uses.

A linked relay is not only file access. From the webchat, operators can open desktop sessions, terminals, VS Code/code-server, context editors, memory editors, and provider runtime views while the agent works.

Desktop RelayAgents can inspect and control a full desktop through screen tools. With a local relay started with `allow_local`, that desktop can be the local machine where the relay runs; otherwise it is the relay container or configured remote desktop.

Terminals on the relayThe webchat can open terminals for the relay Docker workspace, the host-local relay surface when allowed, and the relay server/runtime shell used for diagnostics.

Interactive provider tmuxCLI-backed interactive providers such as Claude Code interactive and Antigravity/Agy run in tmux-backed sessions, so their live terminal state can be observed and debugged instead of being a black box.

Desktop access

Screen, browser, VNC, and full desktop tools let agents operate real UI workflows when APIs are not enough.

Relay terminals

Open shell sessions against the remote Docker relay, the local host relay when `allow_local` is enabled, or the relay server runtime.

Context editor

Inspect and adjust the conversation context that agents will actually see before a turn runs.

Memory editor

Review, correct, or remove durable memories so long-term context stays useful and auditable.

VS Code/code-server

Open a browser editor attached to the relay workspace for hands-on review while agents continue in the same conversation.

Provider sessions

Watch tmux-backed CLI providers and relay diagnostics when a provider, login, or long-running tool needs direct inspection.

What you can do

From first chat to operational workflows.

Agentic coding

Run coding agents against a linked workspace with relay-backed read, edit, shell, grep, browser, and project graph tools.

Multi-agent work

Delegate tasks, assign plans, verify outputs, and run different providers in the same conversation.

Desktop and browser tasks

Operate controlled desktops, browsers, VNC sessions, and forwarded local services from the same runtime.

Media generation

Generate and transform images, video, audio, voice, 3D assets, try-on outputs, and upscaled media into FileStore.

Deterministic automations

Run NiFi-style JSON flows with 100+ task types across IO, data, control, system, and AI categories.

Self-hosted security model

Use explicit relays, stored secrets, provider boundaries, private gateways, and per-agent permissions.

Agent-created deterministic flow pattern — Agents are useful at design time; flows keep the schedule reliable at runtime.

The practical difference

Not every workflow should keep an agent in the loop forever.

For recurring jobs, PawFlow lets the agent design or maintain the automation, then the flow engine runs the explicit graph. You decide where LLM calls are allowed and where execution must stay deterministic.

Agent modeExplore, code, inspect, reason, delegate.

Flow modeSchedule, route, retry, transform, deliver.

Five-minute path

Install it, open the wizard, start the first conversation.

The Docker installer checks prerequisites, starts PawFlow, opens the bootstrap wizard, creates the admin user, configures the first LLM service, and deploys the starter agent flow.

Open quickstart Browse docs