FlowAgent is a Chrome Manifest V3 extension for chat-first browser workflows. You describe a task in plain language, FlowAgent turns it into a bounded step plan, keeps approvals in the chat thread, and runs the workflow inside the browser with your own model keys.
- A sidepanel-first browser agent for structured web workflows
- Local-first and BYOK
- Chat UI for planning, approval, and execution updates
- Separate memory window so the main thread stays clean
FlowAgent is focused on browser workflows, not general desktop automation.
Supported capabilities today:
- plan a workflow from natural language
- edit workflow steps inline
- approve, dismiss, run, pause, resume, or stop
- gate risky actions behind explicit approval
- execute bounded browser actions across tabs
- save reusable workflows and selector memory locally
- use multiple provider accounts and switch models from the top bar
sidepanel.html / sidepanel.js / sidepanel.css- Main chat-first interface
options.html / options.js / options.css- Separate memory window for saved workflows, notes, and domain memory
src/background/service-worker.js- Extension entrypoint and message router
src/background/agent-engine.js- Planner/executor orchestration
src/content/page-tools.js- Bounded browser actions injected on demand
Supported workflow step kinds:
open_urlclicktypeselect_optionwait_forextract_textextract_listscrollswitch_tabclose_tabsummarizeask_user
Each step is structured JSON. FlowAgent does not execute arbitrary model-generated code.
- Full plan review before execution
- Risky browser actions require explicit user approval
- Site permissions are requested at runtime
- Memory stays in
chrome.storage.local
FlowAgent currently supports user-managed keys for:
- OpenAI
- Google Gemini
- Claude
- NVIDIA
The model picker only shows text LLMs relevant to planning and execution.
npm test
npm run buildBuild output:
- unpacked extension:
dist/flowagent - zip package:
dist/flowagent.zip
The current direction is workflow-first browser automation:
- better planning
- better approval UX
- better memory
- stronger reusable workflows
Recorder, cloud sync, and non-browser automation are intentionally out of scope for now.