Pie · Open-Source AI Agent for Browser & Tabs

Tracking since May 9, 2026.

Changelog

• Jun 1, 2026
  description

  Pie is an open-source AI browser agent for Chrome. Type a task in
  natural language — "summarize this page", "close inactive tabs",
  "fill out this form using my saved data" — and Pie's AI agent plans
  the steps and executes them across your browser tabs.
  
  ✨ What makes Pie different:
  
  🔑 BYOK (BRING YOUR OWN KEY) — No subscription, no Pie backend, no
  telemetry, no proxy. Your API key is encrypted locally with AES-GCM
  and only ever sent as the Authorization header on direct calls to
  your chosen provider.
  
  🤖 NATIVE TOOL CALLING — Not a prompt-glued wrapper. Pie uses
  Anthropic tool_use blocks and OpenAI function_calling to drive a
  typed tool registry: DOM actions, cross-tab orchestration, page
  content extraction.
  
  🎯 SKILLS — Save any prompt as a reusable skill with a scoped tool
  whitelist. Run with /skill_name. The AI agent can author its own
  skills too, so workflows the model just figured out can be replayed
  next session.
  
  ⌨️ CANVAS EDITOR SUPPORT — Pie supports Chrome DevTools Protocol
  keyboard injection (opt-in), so it works with canvas-based editors
  like Google Docs, Notion, and Lark Docs where standard DOM events
  fail.
  
  📑 MULTI-SESSION — Conversations survive Service Worker restarts.
  Archived sessions evict on storage pressure (LRU + 30-day delete).
  Pin tabs per session so Pie's agent stays scoped to your task.
  
  🌐 SIDE PANEL, NOT POPUP — Pie lives in Chrome's side panel and
  stays open while you browse. Chat, run AI tasks, manage tabs
  without losing context.
  
  🔒 PRIVACY-FIRST — No Pie servers. No analytics. No third parties.
  Page content sent to the LLM is wrapped in <untrusted_*> markers to
  defeat prompt-injection attempts from page DOM. Apache 2.0 licensed
  — source at github.com/WiseriaAI/pie-ai-agent.
  
  ⚙️ How to use:
  1. Install Pie
  2. Open the side panel, go to Settings, paste your API key
  3. Switch to Chat and tell Pie what to do
  
  Requires Chrome 114+ or any Chromium browser with side-panel
  support (Edge, Brave, Arc).
  
  Pie is open-source software. Found a bug? Want a feature? Issues
  and PRs welcome at github.com/WiseriaAI/pie-ai-agent.

  Pie is an open-source AI browser agent for Chrome. Type a task in
  natural language — "summarize this page", "close inactive tabs",
  "fill out this form using my saved data" — and Pie's AI agent plans
  the steps and executes them across your browser tabs.
  
  ✨ What makes Pie different:
  
  🔑 BYOK (BRING YOUR OWN KEY) — No subscription, no Pie backend, no
  telemetry, no proxy. Your API key is encrypted locally with AES-GCM
  and only ever sent as the Authorization header on direct calls to
  your chosen provider.
  
  🤖 NATIVE TOOL CALLING — Not a prompt-glued wrapper. Pie uses
  Anthropic tool_use blocks and OpenAI function_calling to drive a
  typed tool registry: DOM actions, cross-tab orchestration, page
  content extraction.
  
  🎯 SKILLS — Save any prompt as a reusable skill with a scoped tool
  whitelist. Run with /skill_name. The AI agent can author its own
  skills too, so workflows the model just figured out can be replayed
  next session.
  
  ⌨️ CANVAS EDITOR SUPPORT — Pie supports Chrome DevTools Protocol
  keyboard injection (opt-in), so it works with canvas-based editors
  like Google Docs, Notion, and Lark Docs where standard DOM events
  fail.
  
  🌐 WEB SEARCH — Pie's agent can search the web for current info via
  Tavily (BYOK — bring your own Tavily key, 1,000 free queries / month).
  Search results are sandboxed in untrusted markers and the agent
  auto-drills into the top URLs as pinned tabs, then synthesizes with
  real citations. Knowledge questions, fact-checks, "research X for me"
  — Pie figures out when to search and when to use the tab you're on.
  
  📑 MULTI-SESSION — Conversations survive Service Worker restarts.
  Archived sessions evict on storage pressure (LRU + 30-day delete).
  Pin tabs per session so Pie's agent stays scoped to your task.
  
  🌐 SIDE PANEL, NOT POPUP — Pie lives in Chrome's side panel and
  stays open while you browse. Chat, run AI tasks, manage tabs
  without losing context.
  
  🔒 PRIVACY-FIRST — No Pie servers. No analytics. No third parties.
  Page content sent to the LLM is wrapped in <untrusted_*> markers to
  defeat prompt-injection attempts from page DOM. Apache 2.0 licensed
  — source at github.com/WiseriaAI/pie-ai-agent.
  
  ⚙️ How to use:
  1. Install Pie
  2. Open the side panel, go to Settings, paste your API key
  3. Switch to Chat and tell Pie what to do
  
  Requires Chrome 114+ or any Chromium browser with side-panel
  support (Edge, Brave, Arc).
  
  Pie is open-source software. Found a bug? Want a feature? Issues
  and PRs welcome at: github.com/WiseriaAI/pie-ai-agent.

• Jun 1, 2026
  host_permissions

  <all_urls>, https://api.anthropic.com/*, https://api.openai.com/*, https://openrouter.ai/*, https://api.minimax.chat/*, https://open.bigmodel.cn/*, https://dashscope.aliyuncs.com/*, https://generativelanguage.googleapis.com/*, https://api.deepseek.com/*

  <all_urls>, https://api.anthropic.com/*, https://api.openai.com/*, https://openrouter.ai/*, https://api.minimax.chat/*, https://open.bigmodel.cn/*, https://dashscope.aliyuncs.com/*, https://generativelanguage.googleapis.com/*, https://api.deepseek.com/*, https://api.xiaomimimo.com/*, https://api.tavily.com/*

• Jun 1, 2026
  permissions

  activeTab, sidePanel, storage, tabs, tabGroups, scripting, debugger, webNavigation

  activeTab, sidePanel, storage, tabs, tabGroups, scripting, debugger, webNavigation, offscreen, downloads

• May 25, 2026
  description

  Bring your own LLM API key. Pie reads, summarizes, and automates pages in a side panel — local keys, no servers, full consent.
  
    Detailed description
  
    Pie is an AI assistant that lives in Chrome's side panel. It reads the page you're on, automates multi-step browsing tasks, and helps you manage tabs — using your own LLM API key. There is no Pie account, no developer-operated server, and no subscription.
  
    WHAT PIE CAN DO
  
    • Chat with awareness of the current page. Ask "summarize this article", "explain this code block", "translate this section" — Pie extracts the visible text (with credential-looking inputs scrubbed out) and sends only that to your LLM.
  
    • Run multi-step agent tasks. Describe what you want in natural language; the model plans steps and executes them through Pie's tool registry — clicking, typing, selecting, scrolling, taking structured snapshots of interactive elements, switching tabs.
  
    • Manage your tabs. Built-in skills auto-group tabs by topic, close duplicates, or close long-inactive tabs. Seven cross-tab tools (list, activate, close, group, ungroup, move, fetch readable content from another tab) are also available individually.
  
    • Save your own skills. A skill is a saved prompt template with an explicit tool whitelist. Run any skill from chat by typing /skill_name. The agent itself can author skills, too — useful for capturing a workflow you just walked through so it's reusable in the next session.
  
    • Type into canvas editors (opt-in). For Lark Docs, Google Docs, and other editors that don't honor standard DOM input events, Pie can simulate keyboard input via the Chrome DevTools Protocol. This feature is OFF BY DEFAULT and must be explicitly enabled in Settings before any debugger session is attached.
  
    ASKS BEFORE IT ACTS
  
    Pie classifies every tool call before it runs. Anything irreversible or cross-origin requires you to approve a confirmation card first. The card shows you:
  
    • The exact tool name and the raw argument the agent intends to pass — so you can spot a mistake or a prompt-injected instruction in page DOM before it runs.
    • The origin and tab title for each affected tab, called out separately when an action would cross the original task's pinned origin.
    • Approve, Reject, or Discard buttons (Discard abandons the agent's plan entirely).
  
    If you reject three actions on the same task, Pie terminates the task automatically — rather than burning your tokens on a plan you've already disagreed with.
  
    SUPPORTED PROVIDERS
  
    • Anthropic Claude — native API, native tool_use blocks
    • OpenAI — function_calling
    • OpenRouter — OpenAI-compatible
    • MiniMax — OpenAI-compatible
    • ZhiPu (智谱) — OpenAI-compatible
    • Bailian (阿里云百炼) — OpenAI-compatible
  
    Adding a provider is a registry entry plus a host permission. Gemini and local Ollama support are on the public roadmap.
  
    PRIVACY AT A GLANCE
  
    • Your API key never leaves your device, except as an Authorization header on the direct provider call.
    • Page content is sent only to the LLM provider you configured, only as part of an explicit task you triggered.
    • Pie has no developer-operated server. No telemetry. No analytics. No third parties.
    • All your data — sessions, skills, preferences — lives in chrome.storage.local. Uninstalling Pie deletes all of it.

  Pie is an open-source AI browser agent for Chrome. Type a task in
  natural language — "summarize this page", "close inactive tabs",
  "fill out this form using my saved data" — and Pie's AI agent plans
  the steps and executes them across your browser tabs.
  
  ✨ What makes Pie different:
  
  🔑 BYOK (BRING YOUR OWN KEY) — No subscription, no Pie backend, no
  telemetry, no proxy. Your API key is encrypted locally with AES-GCM
  and only ever sent as the Authorization header on direct calls to
  your chosen provider.
  
  🤖 NATIVE TOOL CALLING — Not a prompt-glued wrapper. Pie uses
  Anthropic tool_use blocks and OpenAI function_calling to drive a
  typed tool registry: DOM actions, cross-tab orchestration, page
  content extraction.
  
  🎯 SKILLS — Save any prompt as a reusable skill with a scoped tool
  whitelist. Run with /skill_name. The AI agent can author its own
  skills too, so workflows the model just figured out can be replayed
  next session.
  
  ⌨️ CANVAS EDITOR SUPPORT — Pie supports Chrome DevTools Protocol
  keyboard injection (opt-in), so it works with canvas-based editors
  like Google Docs, Notion, and Lark Docs where standard DOM events
  fail.
  
  📑 MULTI-SESSION — Conversations survive Service Worker restarts.
  Archived sessions evict on storage pressure (LRU + 30-day delete).
  Pin tabs per session so Pie's agent stays scoped to your task.
  
  🌐 SIDE PANEL, NOT POPUP — Pie lives in Chrome's side panel and
  stays open while you browse. Chat, run AI tasks, manage tabs
  without losing context.
  
  🔒 PRIVACY-FIRST — No Pie servers. No analytics. No third parties.
  Page content sent to the LLM is wrapped in <untrusted_*> markers to
  defeat prompt-injection attempts from page DOM. Apache 2.0 licensed
  — source at github.com/WiseriaAI/pie-ai-agent.
  
  ⚙️ How to use:
  1. Install Pie
  2. Open the side panel, go to Settings, paste your API key
  3. Switch to Chat and tell Pie what to do
  
  Requires Chrome 114+ or any Chromium browser with side-panel
  support (Edge, Brave, Arc).
  
  Pie is open-source software. Found a bug? Want a feature? Issues
  and PRs welcome at github.com/WiseriaAI/pie-ai-agent.

• May 25, 2026
  short_description

  BYOK Chrome Extension Agent — bring your own API key for page understanding, multi-step task automation, and smart tab management.

  Open-source AI browser agent. Bring your own Claude / ChatGPT / Gemini key — no backend, no telemetry, no proxy. 8 providers.

• May 25, 2026
  name

  Pie

  Pie · Open-Source AI Agent for Browser & Tabs

• May 25, 2026
  host_permissions

  <all_urls>, https://api.anthropic.com/*, https://api.openai.com/*, https://openrouter.ai/*, https://api.minimax.chat/*, https://open.bigmodel.cn/*, https://dashscope.aliyuncs.com/*

  <all_urls>, https://api.anthropic.com/*, https://api.openai.com/*, https://openrouter.ai/*, https://api.minimax.chat/*, https://open.bigmodel.cn/*, https://dashscope.aliyuncs.com/*, https://generativelanguage.googleapis.com/*, https://api.deepseek.com/*

• May 25, 2026
  permissions

  activeTab, sidePanel, storage, tabs, tabGroups, scripting, debugger

  activeTab, sidePanel, storage, tabs, tabGroups, scripting, debugger, webNavigation

Permissions & access

Permissions

activeTabsidePanelstoragetabstabGroupsscriptingdebuggerwebNavigationoffscreendownloads

Host access

<all_urls>, https://api.anthropic.com/*, https://api.openai.com/*, https://openrouter.ai/*, https://api.minimax.chat/*, https://open.bigmodel.cn/*, https://dashscope.aliyuncs.com/*, https://generativelanguage.googleapis.com/*, https://api.deepseek.com/*, https://api.xiaomimimo.com/*, https://api.tavily.com/*

Screenshots