Self-hosted agent runtime for real infrastructure.
Run durable AI agents against your own files, tools, browsers, desktops, services, and workflows.

Website · Quickstart · Docs · Releases

▶ 70-second demo — a text-only GLM 5.2 operates a Linux desktop, opens Chromium, and plays a song on YouTube. Every screenshot is described to it by a separate vision model (delegated vision). The demo video itself was cut, narrated, and scored by a Claude agent running inside PawFlow.

👉 Screenshots, live feature tour, quickstart, and full documentation live on pawflow.allcolor.org.

The Ask PawFlow help bot on the website is powered by a PawFlow agent flow (web_help_bot, behind /api/help).

PawFlow is the runtime layer between chat agents, local tools, and production workflows. The server keeps conversations, context, memory, files, flows, and provider sessions durable. Relays execute filesystem, shell, browser, desktop, and media tools next to the machines where the work actually happens.

Use it when a hosted coding assistant is too boxed-in, a workflow tool is too rigid, and a library is not enough runtime.

PawFlow gives agents a real operating surface without handing your workspace to a vendor-controlled agent cloud.

• Relay-backed tools: read, edit, grep, run commands, browse, control desktops, generate media, and inspect projects through explicit relay routes.
• Service Tunnels: reach a relay-approved TCP service from another relay through a loopback-only listener, with owner-scoped records, short-lived signed grants, and FRPS Login/NewProxy authorization.
• Purpose-built context: conversations, memory, knowledge graphs, agent diaries, relay-scoped project graphs and wikis, durable todo state, expiring scratchpads, files, and buckets each keep their own scope and lifetime across restarts.
• Skill learning loop: agents crystallize hard-won procedures into skills, update skills that proved wrong during use, and get conservative skill drafts proposed from compaction summaries; skill usage is tracked and a skillCurator flow task produces review-first maintenance reports — nothing is archived or promoted without your confirmation.
• Encryption at rest (opt-in): per-conversation passphrase encryption of message content, thinking, and tool I/O (and conv-scoped relay workspaces via CryFS); keys live in RAM only, so a stopped server leaves only ciphertext on disk. Off by default and transparent to conversations that don't use it.
• External secret providers: keep the existing logical secret names while resolving values from AWS Secrets Manager or SSM, HashiCorp Vault, Azure Key Vault, Google Cloud Secret Manager, or Keeper. Remote values are cached only in memory, and conversation plus per-agent allowlists bound their exposure.
• Multi-provider agents: mix Codex interactive, Claude Code interactive, Antigravity/Agy, Gemini CLI, Anthropic, OpenAI, and OpenAI-compatible services per agent or conversation. The old Codex app-server and Claude Code -p agent transports remain available only for legacy configurations.
• Native CLI engines, not API reimplementations: subscription providers run the real interactive Codex, Claude Code, Antigravity, and Gemini CLI engines per conversation — native harness and reasoning preserved — with native Codex plugins (codex_plugins) and Claude Code plugin marketplaces (claude_plugins/claude_marketplaces) declarable per LLM service.
• External agent interoperability: publish several agents from one conversation as independent authenticated MCP endpoints, publish one or more A2A endpoints, call remote A2A agents, or attach Claude Code, Codex, Agy/Gemini, OpenCode, JCode, Pi, or Hermes as a first-class external MCP agent.
• Delegated vision: pair a strong text-only reasoning model with a separate vision-enabled LLM so uploads, screenshots, and desktop views become detailed descriptions with UI coordinates before the reasoning turn. Images sent to a text-only model are never silently dropped: any model — including free-tier ones — gets vision, and clicks stay accurate because coordinates come from the vision model, verified locally by the pre-click screen guard.
• Shared clients: continue the same conversation from the web UI, PawCode CLI, VS Code, the Android app, API clients, or channel integrations.
• Deterministic flows: turn repeated work into NiFi-style DAGs with scheduling, backpressure, checkpoints, approvals, and explicit LLM steps.
• Package ecosystem: distribute agents, skills, tools, services, flow tasks, flows, and UI extensions as signed .pfp packages or import skills from supported marketplaces.

• Agentic coding sessions against a linked workspace, with persistent context and auditable tool output.
• Multi-agent operations where planners, coders, reviewers, researchers, and verifiers work in the same conversation.
• MCP and A2A gateways that expose an existing PawFlow conversation to external clients or connect remote agents to the same durable runtime.
• Browser and desktop automation for workflows that do not have clean APIs.
• Vision-guided desktop agents built from a text-only reasoning model and an independently selected vision model.
• Realtime voice conversations with your agents — speech-to-speech sessions (OpenAI Realtime or Gemini Live) with live captions, barge-in, tool use, and Telegram voice-note replies, persisted as normal conversation history.
• Media pipelines that create images, video, audio, 3D assets, voice, and FileStore outputs.
• Scheduled operational flows: daily digests, inbox triage, data transforms, reports, monitoring, and webhook-driven automation.
• Reusable packages and registries for sharing internal or community agents, skills, tools, services, flow tasks, flows, and UI extensions.
• Live help bots: the Ask PawFlow assistant on pawflow.allcolor.org is powered by a PawFlow agent flow (web_help_bot) answering questions in real time, with a Telegram counterpart (telegram_help_bot).
• Portable conversations with full PawFlow archives, including optional FileStore attachments and generated files.

📖 Prefer a guided version with screenshots? Follow the quickstart on the website.

The easiest path is the Docker installer from the latest release. It starts PawFlow, opens the bootstrap wizard, creates the first admin user, configures the selected LLM services, deploys the starter flow, and opens your first agent conversation.

Downloadable artifacts are published on the latest GitHub release: installer zip, PawCode packages, Relay CLI archives, Relay Desktop installers, the Android APK (pawflow-android-<version>-debug.apk), checksums, and source archives.

PAWFLOW_VERSION=$(curl -fsSL https://api.github.com/repos/allcolor/PawFlow-Agents/releases/latest \
  | python3 -c 'import json,sys; print(json.load(sys.stdin)["tag_name"])')

curl -L -o "pawflow-install-$PAWFLOW_VERSION.zip" \
  "https://github.com/allcolor/PawFlow-Agents/releases/download/$PAWFLOW_VERSION/pawflow-install-$PAWFLOW_VERSION.zip"
unzip "pawflow-install-$PAWFLOW_VERSION.zip"
cd "pawflow-install-$PAWFLOW_VERSION"

bash scripts/install-pawflow.sh --port PORT --pull-images

--version is optional: when omitted the installer resolves the latest published release from GitHub. Pass --version "$PAWFLOW_VERSION" to pin a specific release.

On Windows PowerShell with Docker Desktop Linux containers, use the bundled PowerShell installer instead:

powershell -ExecutionPolicy Bypass -File scripts/install-pawflow.ps1 -Port PORT -PullImages

-Version is likewise optional and defaults to the latest published release; pass -Version $env:PAWFLOW_VERSION to pin a specific release.

Check and apply release updates with:

bash scripts/install-pawflow.sh --check-updates
bash scripts/install-pawflow.sh --self-update
bash scripts/install-pawflow.sh --version NEW_VERSION --port PORT --pull-images

The update command recreates the PawFlow server container on the requested image while keeping persistent data under PAWFLOW_HOME, then removes older PawFlow server/relay image tags unless --keep-old-images is set.

A running deployment can also update itself from the browser: Admin → Update server runs the same steps in a throw-away pawflow-updater container, then waits for a different server process to answer /health before reloading the page. If it never does, the panel says which failure happened and points at docker logs pawflow-updater. See Docker for the mechanism.

On Linux hosts with AppArmor (Ubuntu, Debian, ...), the installer also loads the PawFlow AppArmor profiles (pawflow-mount for provider pool containers, pawflow-relay for relay containers) into /etc/apparmor.d/ — sudo may prompt once. This confines the containers' mount privileges to exactly what they need; without the profiles PawFlow still works but those containers run apparmor=unconfined. Skip with --skip-apparmor, or load them manually later:

sudo install -m 644 docker/apparmor/pawflow-mount docker/apparmor/pawflow-relay /etc/apparmor.d/
sudo apparmor_parser -r -W /etc/apparmor.d/pawflow-mount /etc/apparmor.d/pawflow-relay

Hosts without AppArmor (Windows/macOS Docker Desktop, WSL2, SELinux distros) are detected and skipped automatically — nothing to do there.

Open the installer at:

https://localhost:PORT/install

The first-run Private Gateway key is RoyBatty. Finalizing the wizard replaces it.

git clone https://github.com/allcolor/PawFlow-Agents.git
cd PawFlow-Agents
pip install -r requirements.txt
python cli.py start --host 0.0.0.0 --port PORT

Open the web chat at:

http://localhost:PORT/chat

The web UI is the main operator surface: chat, context editor, memory editor, file attachments, relay tools, desktop entry points, terminals, provider sessions, and flow actions in one place.

PawCode is a terminal client for the same PawFlow conversations. It can be used interactively or in Claude Code-compatible stream-JSON mode.

pawcode --server http://localhost:PORT

echo '{"type":"user","message":{"role":"user","content":"hello"}}' | \
  pawcode --input-format stream-json --output-format stream-json

The VS Code extension attaches to the same PawFlow conversation and resource panel from inside your editor.

The native Android client (pawflow-android/, APK on every release) manages multiple server profiles — HTTPS origin plus private gateway key, encrypted with Android Keystore — signs in natively (built-in credentials or OAuth2 in a Custom Tab with a PKCE handoff that never exposes the session token), then opens the authenticated webchat in parallel native tabs. Webchat downloads go through the system DownloadManager, and the native chrome folds away to give the chat the whole screen. See docs/ANDROID_APP.md.

Telegram bridges chats into the same conversations: message a BotFather bot and agents reply inline (web login uses the Telegram Login Widget).

┌─────────────────────────────────────────────────────────────────┐
│                        PawFlow Server                           │
│                                                                 │
│  ┌──────────┐  ┌──────────┐  ┌─────────┐  ┌────────────────┐  │
│  │  Agents  │  │ Pipeline │  │   Auth   │  │  Web Chat UI   │  │
│  │  (LLM +  │  │  Engine  │  │ Gateway  │  │  (SSE, files,  │  │
│  │  tools)  │  │ (100+    │  │ (9 OAuth │  │   context,     │  │
│  │          │  │  tasks)  │  │ provid.) │  │   commands)    │  │
│  └────┬─────┘  └──────────┘  └──────────┘  └────────────────┘  │
│       │                                                         │
│  ┌────┴─────────────────────────────────────────────────────┐  │
│  │              90+ Tool Handlers (via relay)                │  │
│  │  bash, read, write, edit, glob, grep, web_search,        │  │
│  │  screen, browser, generate_image, generate_video,        │  │
│  │  generate_audio, generate_3d, clone_voice, speak,        │  │
│  │  remember, kg_add, project_graph, delegate, plans, ...   │  │
│  └──────────────────────────┬───────────────────────────────┘  │
│                             │ WebSocket                        │
└─────────────────────────────┼──────────────────────────────────┘
                              │
                    ┌─────────┴─────────┐
                    │   Relay (Docker)   │  ← runs on user's machine
                    │   or native host   │
                    └───────────────────┘

The server hosts the API, agent orchestration, pipeline engine, and web UI. A relay runs on the user's machine (or in a Docker container) and executes tools — filesystem access, bash commands, code edits — over a WebSocket connection. This means agents can manipulate your local codebase without the server needing direct access to your files. Connect a relay with the relay CLI or Relay Desktop to attach workspaces, desktops, browsers, and terminals.

Switch providers per agent, per conversation, or globally. API keys normally use direct openai/anthropic services; subscription logins use the matching interactive CLI-backed provider (codex-interactive, claude-code-interactive, or antigravity-interactive). claude-code (cc -p) and codex-app-server are legacy agent transports kept for existing configurations; do not select them for new agent services. Self-hosted and third-party LLMs can use the OpenAI-compatible endpoint (base_url override). See LLM Providers.

An llmAggregator consults several direct llmConnection services in parallel before a final LLM answers or performs the requested work. Each advisor inspects the request and returns an internal implementation plan; only the final aggregator streams to the user and runs the normal visible tool loop.

{
  "type": "llmAggregator",
  "aggregator_llm_service": "llm_final",
  "advisor_llm_services": ["llm_architect", "llm_reviewer"],
  "max_parallel_advisors": 2,
  "advisor_max_iterations": 20,
  "failure_policy": "best_effort",
  "enforce_read_only": true
}

Advisor contexts are silent and ephemeral. With enforce_read_only: true (the default), advisors receive a fail-closed read-only tool set, including through CLI-backed providers; the final LLM keeps the conversation's normal tools and approval policy. Advisor usage is tracked separately so it does not inflate the main context gauge. See the multi-LLM aggregator how-to and technical guide.

An llmRouter selects one direct llmConnection for each logical agent turn. It supports ordered, round_robin, sticky_round_robin, and least_recently_used selection. The immutable candidate plan remains fixed for every later LLM/tool iteration in that turn; a classified provider failure advances within that snapshot.

{
  "type": "llmRouter",
  "strategy": "sticky_round_robin",
  "candidates": [
    {"service_id": "llm_primary", "priority": 10, "enabled": true},
    {"service_id": "llm_backup", "priority": 20, "enabled": true}
  ]
}

During handoff PawFlow flushes persisted work and cold-starts the next provider from the current context. Completed work is not replayed, unresolved tool outcomes are marked unknown, and cancellation or force stop never affects route health. Health and Explain actions expose sanitized operational state. Legacy llmFailover definitions migrate once to ordered routers; invalid user/conversation definitions are disabled and quarantined, while invalid global definitions stop startup for administrator repair. See the technical service reference.

An LLM service with supports_vision: false can delegate every incoming image to another vision-enabled llmConnection. PawFlow asks that service for visible text, layout, UI controls, states, and approximate pixel coordinates, then replaces the image with that description only for the text model's outbound call. The stored conversation retains the original image.

{
  "default_model": "glm-5.2:cloud",
  "supports_vision": false,
  "vision_llm_service": "ollama_gemma4_vision"
}

This lets a model such as GLM 5.2 inspect uploads and use screen/see/read results through Gemma 4 Cloud, while GLM remains the agent's reasoning model and desktop-tool caller. Descriptions are cached by image content hash. See the GLM 5.2 + Gemma 4 how-to and the technical provider reference.

Agents have persistent cognition plus scoped work state:

Memory and diary digests plus active todo state are injected into turn context. Scratchpad bodies are deliberately pull-only: the agent sees a compact topic/count hint and calls scratchpad to retrieve relevant notes. See Cognitive Tools for the memory/KG/diary/todo/scratchpad decision guide.

The full Agent Tool Selection guide also distinguishes file/search/edit tools, artifacts, user questions and notifications, cognitive stores, resources, packages, skills, tasks, and flows.

Agents receive a compact ## Tool selection map filtered to their actual tool registry. They can request a complete comparison on demand with get_tool_schema(family="delegation"), then inspect exact parameters with get_tool_schema(tool_name="delegate"). The full Markdown guide is not copied into every prompt.

• Delegate tasks to sub-agents with delegate()
• Each sub-agent gets its own LLM, tools, and conversation context
• Agents can run in parallel or sequentially
• Git worktree isolation for parallel coding tasks is on the roadmap

• Create structured multi-step plans with create_plan()
• Step-by-step execution with approval gates
• Assign steps to different agents
• Verify completed work before moving on

A PawFlow secret name can store its encrypted value locally or point to a read-only entry in AWS Secrets Manager, AWS SSM Parameter Store, HashiCorp Vault KV, Azure Key Vault, Google Cloud Secret Manager, or Keeper Secrets Manager. Expressions, flows, services, packages, and tools keep using the same logical name, so moving a value out of PawFlow does not rewrite consumers.

External values are materialized through a bounded in-memory TTL cache. Optional conversation and per-agent allowlists intersect, so an agent can never expand the conversation's secret envelope. Resolution fails closed and never falls back to a lower-scope value when the winning external reference is denied or unavailable. See External Secret Providers.

Publish one or more attached agents from an existing conversation as independent authenticated Streamable HTTP MCP endpoints. Each publication has its own endpoint, keys, tool allowlist, client lease, and terminal registration. Claude Code, Codex, Gemini CLI/Agy, OpenCode, JCode, Pi, Hermes, and other MCP clients can then use the selected agent's PawFlow tools under the owner's normal permissions, hooks, and relay configuration.

The optional local stdio bridge also shares the CLI's current project directory without changing the conversation's default relay. Release assets include a universal ZIP and tar.gz with guided installers for Windows, Linux, and macOS. The wizard configures Claude Code, Codex, Agy, OpenCode, JCode, Pi, and Hermes while keeping API and gateway keys in one private local profile. See the MCP client installation guide and Published Conversation MCP Servers.

An authenticated MCP client can also become a first-class external_mcp conversation agent. Its terminal receives user, delegate, and shared-context A2A turns while PawFlow keeps the conversation, permissions, relay tools, and result routing durable.

Publish one or more conversation agents as authenticated A2A 1.0 HTTP+JSON endpoints, delegate to agents in other PawFlow conversations, or call a generic remote A2A agent with the built-in a2a tool. Resources → A2A provides guided publication, one-time keys, isolated/shared context policy, Agent Card copying, and named local or remote targets. See A2A Integration.

100+ tasks across 5 categories for data processing workflows:

Flows are defined in JSON, executed as DAGs, and support backpressure, checkpointing, crash recovery, parameter contexts, subflows, and CRON scheduling.

PawFlow Packages (.pfp) are signed zip artifacts for distributing PawFlow resources. A package can include agents, prompts, skills, themes, task definitions, flows, service definitions, tools, service providers, flow tasks, task providers, and UI extensions. Install is review-first: PawFlow verifies the package signature and lock file, shows a selectable install plan, records per-object provenance, and executes code-bearing objects through a relay runtime instead of importing third-party code into the server process.

Common package workflows:

/pfp key-create
/pfp build ./my-package.pfpdir --key-env PAWFLOW_PFP_SIGNING_KEY
/pfp inspect ./dist/my-package-1.0.0.pfp
/pfp install ./dist/my-package-1.0.0.pfp --include skill:x,service_provider:y
/pfp dev-load ./my-package.pfpdir --include service_provider:image --secret api_key=my_provider_key
/pfp export --package my.bundle --version 0.1.0 --include agent:helper,flow:daily --out ./my.bundle.pfpdir

Marketplace and registry support is decentralized. Users can add static package registries, search them, inspect remote packages with explicit download confirmation, then install or update selected objects. Skill marketplace import is also supported for Codex/OpenAI skills, Claude/Anthropic plugin marketplaces, HermesHub, and OpenClaw GitHub tree URLs; imports are bounded, reviewed, and never grant tool permissions automatically.

See PawFlow Packages, PFP Developer Guide, PFP Publisher Guide, and Marketplace.

40+ chainable operations for dynamic configuration:

${name:upper}                                     → "ALICE"
${api_key:default("not-set")}                      → uses fallback if empty
${status:equals("active"):then("ON"):else("OFF")}  → conditional logic
${csv_line:split(","):index(0):trim}               → first CSV field, trimmed
${response:json_get("data.items.0.id")}             → extract from JSON
${content:hash_sha256}                              → hash a value
${:uuid}                                            → generate a UUID
${:now:format("yyyy-MM-dd")}                         → "2026-04-08"

Expressions resolve through a cascade: secrets → flow parameters → conversation → user → global → environment variables. See Expression Language docs for the full reference.

• Real-time streaming via SSE
• Two conversation views: Simplified (the default) shows each turn as your message, one live activity block, and the turn's last message below it; Classic keeps the flat transcript. Switch per conversation from the View menu.
• New conversations start in auto permission mode; change it per conversation from the permission selector, or with /permission default|approve_edits|read_only|auto (see Security Model).
• Shared conversations across web, PawCode CLI, VS Code, the Android app, API clients, and channel flows
• File explorer with relay filesystem access
• Context editor (view/edit agent context)
• Conversation management with auto-titles
• Drag & drop file attachments and FileStore outputs
• 60+ slash commands (/agent, /memory, /relay, /run, /plan, /desktop, ...)
• Desktop/VNC entry points plus relay-backed screen actions
• Escape key: 1x = graceful interrupt, 2x = force stop
• Multi-agent with agent switching
• Admin → Update server: pull a new release and restart from the browser, without touching the command line

9 OAuth providers out of the box:

Agents, services, and flows are configured via JSON. Parameters cascade: flow → conversation → user → global.

{
  "llm_service": "claude_code_llm_service",
  "summarizer_service": "claude_code_llm_service",
  "permission_mode": "auto",
  "max_iterations": 200
}

See .env.example for environment variables.

pytest tests/ -v    # 7000+ tests across 360+ test files

See ROADMAP.md for the full roadmap.

Key upcoming areas:

• Stabilization and release hardening
• Manual flow editor
• New media service providers
• Git worktree isolation for parallel agents
• Mobile client (PWA)
• MCP elicitation
• x402 payment policies for published endpoints and outbound calls
• Filesystem hooks
• Full AWS-native remote execution mode
• First-class Discord, Slack, and WhatsApp conversation clients

See CONTRIBUTING.md. In short:

1. Fork & clone
2. pip install -r requirements.txt
3. Make changes, run pytest tests/
4. Open a PR

MIT

🌐 pawflow.allcolor.org — website, feature tour, quickstart, and docs.