Persistent Machines of Loving Grace
Always-on agent runtime with configurable skills and workflows that can run coding agents such as Claude Code.
Open source. On your computer.
Specification
What aflow runs.
Models, coding agents and tools it connects to, and the platform features around them.
- Model providers
- Anthropic
- OpenAI
- xAI
- Fireworks AI
- OpenRouter
Claude, GPT, Gemini, Grok, DeepSeek, Kimi, GLM, Qwen, MiniMax and Mistral for text; Gemini, gpt-image, Veo, Sora and Kling for images and video. Keys are your own.
- Coding agents
- Claude Code
- OpenCode
- CodexSoon
Each run works in its own git worktree on your machine and returns a diff. Nothing is committed or pushed without approval.
- Tools
- MCP servers
- HTTP APIs
- OpenAPI import
- GitHub
- Slack
- Gmail
- Google Calendar
- Google Sheets
- Notion
- Linear
- Jira
- Airtable
- Stripe
- Sentry
- Vercel
- Cloudflare
- Brave Search
Bundled connectors, plus any local or remote MCP server and any HTTP endpoint.
- Runs on
- macOS
- Linux
- Docker
One Compose file, amd64 and arm64 images. Listens on 127.0.0.1 only.
Skills and workflows
Reusable procedures built from steps: model calls, tool calls, code, coding agents and approvals.
Schedules and triggers
Cron, one-off and on-completion schedules, and incoming webhooks.
Approvals
Writes to outside systems wait in the Action Center until approved or denied.
Sandboxed code
Python, Node, Deno and Bash in Docker containers with no network.
Your machine
Connected folders, commands and local MCP servers, confined by an OS sandbox.
Memory
A file-like store per space with vector search on pgvector.
Applets
Small interactive apps a skill can generate, such as dashboards, forms and trackers, each keeping its own saved state.
Evals
Datasets, labelled cases and replayed batches to measure a skill over time.
MCP server
Runs, operations and the catalog, exposed to other agents over MCP.
Single owner, local only. No telemetry. Licensed AGPL-3.0.
The local edition
Start here.
Start with a model key.
Add connections as you go.
Install aflow.
Clone and build it. Docker is all it needs; the first build takes a few minutes.
aflowai/aflow on GitHubgit clone https://github.com/aflowai/aflow.git cd aflow docker compose -f docker-compose.local.yml up -d --buildThen open 127.0.0.1:3001
Or let your coding agent set it up:
Install Aflow from https://github.com/aflowai/aflow by following its README, start it, and open it for me.Connect a model.
Use a key from OpenAI, Anthropic, Google, xAI, OpenRouter or Fireworks AI.
Model usage is billed by your provider.Begin with a conversation.
Ask for something in chat, or turn it into a skill that runs on its own — calling your tools, running code, and handing coding work to Claude Code or OpenCode.
Changes wait for your approval. Every run can be inspected.
Background work continues while aflow and your computer are running.