InnerAgents

InnerAgents turns your browser sidebar into a powerful AI agent workspace. Connect your own API keys from Anthropic, OpenAI, OpenRouter, or any custom provider — your data never goes through a third-party server. MULTI-PROVIDER LLM SUPPORT Use Claude, GPT-4o, Gemini, Mistral, or any OpenAI-compatible model. Switch models per agent. Pricing is displayed in the model selector so you always know what a conversation costs. REAL BROWSER CONTEXT Your agents see what you see. Inject the current page content, a screenshot, or a text selection directly into your conversation. Define URL rules to auto-inject context on specific sites. MCP SERVERS Connect any remote MCP server (SSE or Streamable HTTP). Supports Bearer token and OAuth 2.0 authentication. Your agents gain access to real-world tools — search, calendars, databases, APIs — without writing a single line of code. CUSTOM SKILLS Write JavaScript skills that run sandboxed in the browser. Skills can call Chrome APIs (tabs, scripting, capture) with explicit user consent. Combine skills to automate anything you can do in a browser. NATIVE BRIDGE (OPTIONAL) Install the free InnerAgents Bridge companion app (macOS, Windows, Linux) to give agents filesystem read/write and shell command access, scoped to folders you explicitly allowlist. Every write, delete, and shell command requires your confirmation — there's no "always allow" for shell execution. The extension works fully without it. LLM WIKI & SECOND BRAIN Each agent can keep a personal, local note-taking wiki — PARA folders, daily notes, [[wiki-links]], and backlinks, all searchable and editable from the chat. A private "Second Brain" wiki per agent lets it build up context over time, entirely stored on your device. SESSION RECORDING Record yourself performing a task in the browser — clicks, typed text, page navigation — then replay it deterministically, or turn it into a reusable Skill. Sensitive fields like passwords are never recorded. SECURE CREDENTIALS STORE Save API keys, tokens, and OAuth 2.0 credentials once, then reference them by name from any skill, MCP server, or agent — the actual secret value is never exposed to the model, only its name and description. Supports Basic, Bearer, API key, OAuth 2.0 (with automatic refresh), and native Google auth. STRUCTURED DATA REVIEW Agents can generate a structured review form on the fly — from a JSON Schema — so you can validate, correct, or approve data (extracted records, generated content, parsed results) before it's used, instead of having to parse it out of a chat message. RELIABILITY GUARDRAILS Built-in safeguards keep agents predictable: explicit confirmation before any destructive action (deleting a skill, removing a memory…), tool arguments validated against their schema before execution, automatic retry on transient LLM errors, and automatic MCP reconnection when a server is updated or restarted. AUTOMATION HOOKS Wire custom logic into agent lifecycle events — before/after a message, on tool calls, on skill activation — without modifying the core extension. PERSISTENT MEMORY Agents remember facts across conversations. A global shared memory lets all your agents benefit from the same user context. Memory is stored locally — never in the cloud. MULTI-AGENT COLLABORATION Delegate subtasks to specialized agents. A research agent can hand off to a writing agent, which can hand off to a publishing agent. Trace the full delegation chain in the debug panel. CONTEXT INSPECTOR See exactly what's being sent to the model: system prompt, conversation summary, messages — with token counts and estimated cost per call. OPTIONAL SYNC BETWEEN YOUR DEVICES Sign in with Google to keep your agents, settings, conversations and wiki notes in sync across your devices. Everything is stored in a private, hidden application folder in YOUR Google Drive — InnerAgents has no server holding your data, and it cannot see any of your other Drive files. Turn on end-to-end encryption to make the synced data unreadable to anyone but you. You can erase it all from Settings at any time. Sync is entirely optional: without it, everything stays on your device. RUNS THAT SURVIVE, AND NOTIFICATIONS A long run is never lost: if you close the panel or reload mid-way, InnerAgents offers to resume it where it stopped, and asks again before replaying anything destructive. Optionally, get a Chrome notification and a badge on the icon when an agent finishes, fails or needs your answer while you're elsewhere — one click brings you back to the right conversation. The notification permission is only requested when you switch this on. WEBMCP-READY Pages that expose tools through WebMCP (document.modelContext) are detected automatically, and their tools become available to your agent for that page. THREE LANGUAGES, ONE THEME THAT FOLLOWS YOU The interface is available in English, French and Spanish, and the new Auto theme follows your system's light or dark mode. IMPORT / EXPORT Share agent configurations as self-contained bundles. Import from the InnerAgents Marketplace or from a file, with conflict resolution built in. --- InnerAgents is open and local-first. You bring your own API keys. By default nothing is stored on external servers and your conversations stay on your machine; if you turn on sync, your data goes only to a private folder in your own Google Drive. Get started: add a provider, create an agent, and start chatting.