Crab-Agent

A Chrome extension that turns a large language model into a hands-on browser agent. The agent observes the current page, plans a sequence of actions, executes them through Chrome DevTools Protocol, and repeats until the task is complete or the user takes over.

This repository (crab-ts) is the source code. The user facing distribution lives at crab-agent-extension.

Overview

Crab-Agent is a Chrome MV3 extension. The user opens a side panel, types a task in natural language, and the agent executes it autonomously. The agent can navigate pages, click elements, fill forms, read page content, open and close tabs, upload files, run JavaScript, generate documents, record and replay workflows, schedule future tasks, and maintain persistent memory across sessions.

The execution model is a tool-use loop. The LLM receives the current conversation (including screenshots and page context), selects a tool, the extension executes it, and the result is appended to the conversation for the next LLM call. This continues until the agent calls done.

Architecture

┌─────────────────────────────────────────────────────┐
│  Side Panel (React)                                 │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐  │
│  │ ChatPanel│ │Workflows │ │ Schedule │ │Settings│  │
│  └──────────┘ └──────────┘ └──────────┘ └────────┘  │
│  Zustand stores: taskStore, uiStore, settingsStore, │
│                  workflowStore, memoryStore         │
│  useBgMessage hook via chrome.runtime.Port          │
└────────────────────┬────────────────────────────────┘
                     │  Port: 'side-panel'
                     │  postMessage / onMessage
┌────────────────────▼────────────────────────────────┐
│  Background Service Worker (src/background/index.ts)│
│  Receives new_task / follow_up_task / cancel etc.   │
│  Manages tab group sessions                         │
│  Triggers memory dream cycles                       │
│  Runs chrome.alarms for scheduled tasks             │
│  Calls agent-loop.handleNewTask()                   │
└────────────────────┬────────────────────────────────┘
                     │
┌────────────────────▼────────────────────────────────┐
│  Agent Loop (src/core/agent-loop.ts)                │
│  ┌──────────────────────────────────────────────┐   │
│  │  1. Build system prompt                      │   │
│  │  2. Take initial screenshot                  │   │
│  │  3. Run read_page for element refs           │   │
│  │  4. Loop:                                    │   │
│  │     callLLM -> tool_use response             │   │
│  │     executeTool(name, params)                │   │
│  │     append tool_result                       │   │
│  │     repeat until done()                      │   │
│  └──────────────────────────────────────────────┘   │
│  MessageManager handles compaction and history      │
└────────────────────┬────────────────────────────────┘
                     │
          ┌──────────┴──────────┐
          │                     │
┌─────────▼──────┐   ┌─────────▼──────────────┐
│ LLM Client     │   │ Tool Executors          │
│ llm-client.ts  │   │ src/tools/*.ts          │
│                │   │                         │
│ Anthropic      │   │ CDP: screenshots,       │
│ OpenAI         │   │      clicks, typing     │
│ Google Gemini  │   │ Browser: tabs, nav      │
│ OpenRouter     │   │ Page: read, find, JS    │
│ Ollama         │   │ Files: upload, download │
│ OpenAI-compat  │   │ Docs, GIF, workflows    │
│ Codex OAuth    │   │ Canvas, code editor     │
└────────────────┘   └─────────────────────────┘

Component Map

Background (`src/background/index.ts`)

The single service worker that orchestrates the extension.

Responsibility	Detail
Port management	Maintains a persistent port connection to the side panel
`new_task`	Resolves the active tab, creates a tab group, triggers memory dream if due, builds memory context, calls `handleNewTask`
`follow_up_task`	Queues the message if a task is running, otherwise starts a fresh task with restored history
`cancel_task` / `pause_task` / `resume_task`	Delegates to agent-loop controls
Memory hooks	Intercepts `TASK_OK` events to save memory entries automatically
Tab groups	Collapses the group on `ASK_USER` and expands on resume
Workflows	Manages recording, playback, and analysis
Scheduler	Drives `chrome.alarms` for scheduled tasks
Keep-alive	Self-ping loop during execution to avoid service worker termination

Agent Loop (`src/core/agent-loop.ts`)

The core execution engine.

handleNewTask(
  task, settings, images, sendToPanel,
  llmHistory?, preferredTabId?, workflows?,
  isFollowUp?, memoryContext?
)

runExecutor(exec) is the main loop. It takes a screenshot, injects page context, then iterates: call LLM, parse tool use, execute tool, append result, check for done.

Special tool results handled inline:

Tool result	Behavior
`done`	End task or enter monitor mode
`ask_user`	Pause for input
`suggest_rule`	Forward to user for accept or skip
`memory`	Run CRUD on persistent memory directly
`update_plan`	Plan approval flow

The loop also detects stagnation and injects interrupt messages when the agent loops without progress, tracks active tab changes when tools open or switch tabs, and exports conversation to chrome.storage.local after each task for follow-up continuity.

LLM Client (`src/core/llm-client.ts`)

A single entry point callLLM(messages, settings, useVision, toolSchemas, extraOptions) for every provider.

Highlights:

Translates tool schemas to provider native formats (Anthropic input_schema, OpenAI function calling, Google function_declarations).
Auto-detects an Anthropic endpoint when an openai-compatible URL ends in /v1/messages.
Streams responses (SSE) for OpenAI, OpenRouter, openai-compatible, and Codex providers.
On HTTP 400 with content.str error (vision rejected), retries automatically without images.
Strips images from all but the last image-bearing message to avoid proxy rejections from accumulated screenshots.
Quick Mode disables tool schemas and parses compact text commands instead.

Message Manager (`src/core/message-manager.ts`)

Conversation state for a single task.

Method	Purpose
`compactIfNeeded()`	Strip images from old messages (soft limit 15 MB), then drop oldest non-system messages (hard limit 100 KB)
`exportForStorage()`	Remove base64 images, flatten text-only arrays, preserve tool_use and tool_result pairs
`importFromStorage(history, systemPrompt)`	Restore history and normalize stale content shape

Tools (`src/tools/`)

Each tool is a module with name, description, parameters (JSON Schema), and execute(params, context). Tools return a ToolResult with success, error, and optional flags like isDone, isAskUser, isSuggestRule, isMemoryOp.

The executeTool(name, params, context) function in src/tools/index.ts resolves internal tools first, then external tools, and runs execute.

Other core modules

File	Responsibility
`permission-manager.ts`	Intercepts tool execution to enforce user-controlled safety policies
`memory-manager.ts`	Persistent cross-session storage for facts, preferences, and rules
`dream-engine.ts`	LLM-driven consolidation of memory entries
`task-scheduler.ts`	`chrome.alarms`-based scheduling engine
`tab-group-manager.ts`	Chrome Tab Group lifecycle for agent sessions
`quick-mode.ts`	Low-latency compact command execution path
`cdp-manager.ts`	Chrome DevTools Protocol client
`ax-tree-manager.ts`	Accessibility tree builder and ref ID generator
`state-manager.ts`	Stagnation detection and prompt warnings
`codex-auth.ts`	OAuth PKCE flow for ChatGPT/Codex
`codex-usage-tracker.ts`	Rate-limit tracking from Codex API responses

Side Panel (`src/sidepanel/`)

A React application running in the Chrome side panel.

Component	Role
`ChatPanel`	Message thread, input box, image attachment, live action indicator, suggest-rule accept/reject UI
`SettingsPanel`	Provider, model, API key, base URL, permission mode, system prompt override, tab grouping toggle
`WorkflowList`, `WorkflowSaveModal`	Manage recorded workflows
`ScheduledTaskList`	View and cancel scheduled tasks
`MemoryPanel`	View, delete, and trigger dream consolidation for memory entries
`useBgMessage`	Manages the persistent port connection, heartbeat, and dispatches background messages to stores

Zustand stores: taskStore, uiStore, settingsStore, workflowStore, memoryStore, contextRulesStore.

Agent Loop

handleNewTask called
│
├── Normalize user images (ImageItem[] to data URLs)
├── Build system prompt (memory + context rules + viewport + warnings)
├── Restore conversation history from storage (follow-up) or init fresh
│
├── Take initial screenshot via CDP
├── Run read_page(interactive) and inject element refs into first message
│
└── runExecutor loop (max steps configurable)
    │
    ├── getMessages() -> send to callLLM
    ├── LLM returns tool_use (or text in non-native mode)
    ├── Append assistant tool_use block to conversation
    │
    ├── executeTool(name, params, context)
    │   ├── Internal: done -> exit loop (or enter monitor mode)
    │   ├── Internal: ask_user -> pause, send ASK_USER event to panel
    │   ├── Special: suggest_rule -> forward to user, await accept/skip
    │   ├── Special: memory -> execute CRUD on memoryManager directly
    │   └── External: computer, navigate, find, read_page, tabs_*, etc.
    │
    ├── Append tool_result to conversation
    ├── Stagnation check -> inject interrupt message if looping
    ├── Tab tracking -> update exec.tabId if tool changed active tab
    │
    └── Repeat

On done:

Conversation is exported to storage.
TASK_OK event is sent to the panel with the final answer.
Background intercepts TASK_OK to extract and save memory entries and increment the session counter.
The tab group is ungrouped.

Message Flow

Side Panel                  Background              Agent Loop
─────────────────────────────────────────────────────────────────
sendToBackground({type:'new_task'})
───────────────────────────>
                            handleNewTask()
                            ──────────────────────>
                                                    TASK_START
                            <──────────────────────
sendToPanel(TASK_START)
<──────────────────────────
                                                    THINKING /
                                                    ACTION / STEP
                            <──────────────────────
sendToPanel(execution_event)
<──────────────────────────
                                [loop continues]

                                                    TASK_OK /
                                                    TASK_FAIL
                            <──────────────────────
                            save memory entries
                            ungroup tabs
sendToPanel(TASK_OK)
<──────────────────────────

`execution_event` states

State	Meaning
TASK_START	Task execution began
TASK_OK	Task completed successfully
TASK_FAIL	Task failed after max retries
TASK_CANCEL	Task cancelled by user
TASK_PAUSE	Task paused (ask_user)
STEP_START	New LLM call cycle started
STEP_FAIL	LLM call failed (will retry)
ACTION	Agent is executing a tool
THINKING	LLM is generating
PLANNING	Agent is in planning phase
ASK_USER	Agent needs user input
SUGGEST_RULE	Agent suggests a context rule
COMPACTION	Conversation was compacted
MONITOR_START	Monitoring mode entered
MONITORING	Monitor loop tick
MONITOR_WAKE	Monitor condition triggered
MONITOR_REPORT	Monitor result sent
MEMORY_OP	Memory tool called (add/update/delete/list)
DREAM_START	Dream consolidation started
DREAM_DONE	Dream consolidation finished

LLM Providers

Provider	`provider` value	Tool format	Streaming	Notes
Anthropic	`anthropic`	Native `tool_use`	Optional	Default; uses `tool_choice: auto`; thinking adds `interleaved-thinking-2025-05-14` beta header
OpenAI	`openai`	Function calling	Optional	Chat Completions endpoint
Google Gemini	`google`	`function_declarations`	No	Uses `tool_config: function_calling_config.mode = AUTO`
OpenRouter	`openrouter`	Function calling	Optional	Multi-model router; forwards thinking param for `anthropic/*` models
Ollama	`ollama`	None	No	Local server; `/api/chat` endpoint
OpenAI-compatible	`openai-compatible`	Function calling or native `tool_use`	Optional	Auto-detects Anthropic Messages endpoint when URL ends in `/v1/messages`
Codex OAuth	`codex-oauth`	Function calling	Always	ChatGPT sign-in; PKCE flow; Responses API; tracks rate limits

All providers go through the same callLLM interface. Tool schemas are translated per provider. Vision (image) support is available for every provider and is disabled per request if the endpoint rejects image content.

Tool Registry

Tools are registered in src/tools/index.ts and split into external tools (exposed to the LLM) and internal tools (handled directly by the agent loop).

External Tools

Tool	Purpose
`computer`	Mouse, keyboard, and DOM interaction via CDP. Supports click, type, scroll, drag, screenshot, key press, hover. Ref-based targeting resolves live DOM coordinates
`navigate`	Navigate the active tab to a URL, or go back, forward, or Google search
`read_page`	Read the page DOM and return interactive elements with ref IDs and coordinates. Supports iframe traversal
`find`	Find an element by natural language description. Searches the accessibility tree
`form_input`	Fill form fields, select options, toggle checkboxes
`get_page_text`	Extract readable text content from the page
`tabs_context`	List all open tabs with URLs and titles
`tabs_create`	Open a new tab, optionally at a URL
`switch_tab`	Switch the active tab
`close_tab`	Close a tab
`read_console_messages`	Read browser console output
`read_network_requests`	Read recent network requests and responses
`resize_window`	Change the browser viewport dimensions
`update_plan`	Present a multi-step plan to the user for approval before proceeding
`file_upload`	Upload a file via a file input element
`upload_image`	Upload an image file
`gif_creator`	Record and export a GIF or replay of the task
`suggest_rule`	Propose saving a site-specific interaction rule for future sessions
`memory`	Manage persistent memory: `list`, `add`, `update`, `delete`
`shortcuts_list`	List available keyboard shortcuts on the current page
`shortcuts_execute`	Execute a keyboard shortcut by name
`javascript_tool`	Execute arbitrary JavaScript in the page context
`canvas_toolkit`	Canvas and image manipulation helpers (Figma, Miro, Excalidraw)
`code_editor`	Open an in-panel code editor
`document_generator`	Generate DOCX or HTML documents from task output
`set_of_mark`	Visual overlay for element labeling
`visualize`	Render SVG diagrams or HTML charts inline in the chat
`schedule_task`	Schedule a task for future execution via `chrome.alarms`
`download_file`	Trigger a file download
`run_workflow`	Execute a saved recorded workflow

Internal Tools

Tool	Purpose
`done`	Signal task completion. Accepts `finalAnswer` (text), `monitor` (boolean for watch mode), and monitoring parameters
`ask_user`	Pause execution and prompt the user for input. The agent resumes when the user replies

Memory System

Persistent memory lets the agent remember user preferences, personal information, and project conventions across sessions.

Storage

Key: crab_memory in chrome.storage.local
Maximum entries: 60 (pruned by least recently used)

Entry Structure

interface MemoryEntry {
  id: string
  content: string
  type: 'rule' | 'fact' | 'summary'
  domain?: string         // per-domain entry if set, cross-domain if undefined
  createdAt: number
  lastUsed: number
  useCount: number
  source: 'suggest_rule' | 'dream' | 'manual' | 'auto'
}

memory Tool

The agent uses the memory tool to manage its own memory during a task.

Command	Effect
`memory(command="list")`	Returns all entries with IDs
`memory(command="add", content, type)`	Saves new information (deduplicated by content similarity)
`memory(command="update", id, content)`	Corrects an existing entry
`memory(command="delete", id)`	Removes an entry

The system prompt instructs the agent to use this selectively, only for information the user explicitly shares and that is genuinely useful in future sessions.

System Prompt Injection

Before each task, memoryManager.formatForPrompt(currentDomain) produces a ## Memory section injected into the system prompt. Domain-specific entries appear first, followed by general entries, capped at roughly 1200 characters.

Dream Consolidation

After a threshold of sessions (default 5) and elapsed time (default 24 hours), the background triggers a non-blocking dream cycle:

All memory entries are sent to the LLM with instructions to deduplicate, merge near-identical entries, remove overly generic entries, and return a cleaned JSON array (max 50 entries).
On success, memoryManager.replaceEntries(cleaned) replaces the current entries.
The cycle is fire-and-forget. If the LLM call fails or times out (45 seconds), the original entries are kept unchanged.

Permission Model

The permission manager controls what actions the agent can perform on which domains, preventing unintended modifications to sensitive pages.

Modes

Set via AgentSettings.permissionMode:

Mode	Behavior
`ask` (default)	Request user approval before acting on new domains or sensitive pages. Read-only actions are allowed without prompting
`auto`	Approve all actions automatically
`strict`	Require explicit approval for every action

Permission Types

NAVIGATE, READ_PAGE_CONTENT, CLICK, TYPE, UPLOAD_IMAGE, PLAN_APPROVAL, DOMAIN_TRANSITION.

Grant Durations

Duration	Meaning
`once`	Granted for a single tool use (expires immediately after use)
`always`	Granted for the domain until the session ends or the user revokes it

Domain Safety

Sensitive domains (login pages, financial services, government sites) always require explicit user approval and cannot be granted always.
A blocklist of patterns prevents the agent from accessing known malicious domains.
verifyUrlDomain(tabId, expectedDomain) is called before executing actions to detect unexpected navigation mid-step.

Plan Approval

When the agent calls update_plan, it presents a list of planned domains and actions to the user. If approved, all listed domains receive pre-authorization for the duration of the plan.

Scheduler

The task scheduler uses Chrome's chrome.alarms API to execute tasks at a future time or on a recurring schedule.

Tasks are stored in chrome.storage.local under the key scheduledTasks and alarms are re-registered on service worker startup.
Supports one-time tasks (absolute timestamp or relative delay in seconds) and recurring tasks (5-field cron).
When an alarm fires, the scheduler calls handleNewTask directly (headless if the side panel is not open).
The schedule_task tool allows the agent to schedule follow-up tasks from within a task execution.
The side panel shows a ScheduledTaskList with options to cancel pending tasks.

Workflows

Workflows are recorded sequences of browser interactions that can be replayed on demand.

Recording

The user triggers recording from the side panel.
lib/workflowRecorder.js is injected into the active tab and captures DOM events (clicks, inputs, navigation) as structured steps.
When recording stops, the background receives the action list and the side panel shows a save modal.

Saving

Workflows are saved to chrome.storage.local under crab_workflows.
Each workflow has a name, description, and an optional set of parameterized inputs (such as email address or search query).

Playback

lib/workflowPlayer.js replays steps on the page.
The run_workflow tool lets the agent invoke a saved workflow as part of a larger task.
If the task matches a saved workflow, the agent is instructed to call run_workflow immediately.

Analysis

The background can send a workflow to the LLM for semantic description extraction, which enriches the saved workflow with human-readable step summaries.

Tab Groups

The tab group manager uses the Chrome Tab Groups API to visually group tabs that belong to an agent session.

Behavior	Detail
Session start	The current tab is registered as the session main tab (not grouped yet)
First new tab	A Chrome Tab Group is created with the session name derived from the first words of the task
Subsequent tabs	Added to the same group
Task completion	Tabs are ungrouped
Group title	Reflects the task state: task hint while running, check on completion, X on failure
Collapse behavior	Group collapses on `ask_user` and expands on user reply
Toggle	Settings > Group tabs during tasks

chrome.tabs.onUpdated marks tab context dirty when a session tab navigates, so the agent receives updated page context on the next step.

Quick Mode

Quick Mode is an alternative execution path for lower latency control.

Enabled when settings.quickMode is true.
The LLM receives a compact system prompt and returns single-line commands instead of structured tool calls.
Tool schemas are not sent to the LLM in Quick Mode.

Commands

Command	Action
`CR ref` / `RCR ref` / `DCR ref` / `TCR ref`	Click, right-click, double-click, triple-click by ref ID
`C x y` / `RC x y` / `DC x y` / `TC x y`	Click actions by coordinate
`H x y`	Hover
`T text`	Type text
`K keys`	Press a key or chord (e.g. `K Enter`, `K ctrl+a`)
`S dir amt x y`	Scroll (direction, amount 1-10, position)
`D x1 y1 x2 y2`	Drag from start to end
`N url` / `N back` / `N forward`	Navigate
`J code`	Execute JavaScript
`W`	Wait one second for the page to settle
`ST tabId` / `NT url` / `LT`	Switch tab, new tab, list tabs
`DONE text`	Complete task
`ASK question`	Ask user

Responses end with <<END>> on their own line. The parser extracts an optional <thinking> block and the command lines, then executeQuickModeCommands runs them in sequence.

Configuration

All settings are stored in chrome.storage.local via the settingsStore Zustand store.

Setting	Type	Default	Description
`provider`	string	`anthropic`	LLM provider
`model`	string	(provider default)	Model identifier
`apiKey`	string	`""`	API key for the provider
`baseUrl`	string	(none)	Custom endpoint (Ollama, OpenAI-compatible)
`customModel`	string	(none)	Model name when provider is `openai-compatible` or `ollama`
`maxTokens`	number	(none)	Override max tokens per LLM call
`temperature`	number	(none)	Override temperature
`systemPrompt`	string	(none)	Replace the default system prompt
`permissionMode`	`ask` / `auto` / `strict`	`ask`	Action approval policy
`enableWorkflowRecording`	boolean	true	Show workflow recording controls
`enableScheduledTasks`	boolean	true	Show scheduled task controls
`theme`	`dark` / `light`	(system)	UI theme
`enableMemory`	boolean	true	Enable persistent memory and dream cycles
`enableTabGrouping`	boolean	true	Group agent tabs into a Chrome Tab Group
`codexPlan`	`free` / `plus` / `pro`	(none)	Codex usage tier when provider is `codex-oauth`

Tech Stack

Layer	Technology
UI framework	React 18
Language	TypeScript 5.7
Build	Vite 6 with @crxjs/vite-plugin
Styling	Tailwind CSS and CSS variables
State management	Zustand 5
Markdown rendering	react-markdown, remark-gfm, rehype-highlight
Extension platform	Chrome MV3
Chrome APIs	tabs, tabGroups, storage, debugger, scripting, sidePanel, alarms, webNavigation, downloads, notifications, contextMenus, offscreen, system.display, identity
LLM communication	Fetch (REST) with SSE streaming support

Build and Install

Prerequisites

Node.js 18 or later
Google Chrome 114 or later

Commands

# Install dependencies
npm install

# Development (Vite watch mode for UI iteration)
npm run dev

# Production build (tsc type check, then vite build)
npm run build

The production build outputs to dist/.

Loading into Chrome

Run npm run build.
Open chrome://extensions.
Enable Developer mode.
Click Load unpacked and select the dist/ directory.
Open any web page, then click the extension icon to open the side panel.
Enter your API key in Settings, or pick ChatGPT (Sign in) to authenticate with OAuth.

Chrome Permissions

The extension requests these permissions at install time:

tabs, tabGroups, activeTab, scripting, storage, debugger,
webNavigation, sidePanel, clipboardWrite, clipboardRead,
offscreen, downloads, notifications, system.display,
alarms, contextMenus, identity

Host permissions: <all_urls> (required for CDP-based page interaction and content script injection on any site).

Adding New Features

Adding a new tool

Create src/tools/my-tool.ts exporting myToolSchema and executeMyTool.
Register it in TOOL_REGISTRY and getToolSchemas() in src/tools/index.ts.
Add a case to executeTool() in the same file.
Add a description to prompts/system-prompt.js when the provider does not support native tool-use.

Adding a new LLM provider

Add a case to the switch (provider) block in src/core/llm-client.ts.
Implement _buildXxxRequest().
Implement or reuse _extractXxxResponse() and _extractXxxToolCall().
Add the provider to _isStreamableProvider() if streaming is supported.
Add it to the PROVIDERS list in src/sidepanel/components/settings/SettingsPanel.tsx.
Add models to MODELS_BY_PROVIDER.

Adding a new message type (panel ↔ background)

Add it to the ToBackground or ToSidepanel union in src/sidepanel/lib/types.ts.
Handle it in port.onMessage.addListener in src/background/index.ts.
Send or receive it in useBgMessage or the relevant store.

License

MIT. See LICENSE for details.

Name		Name	Last commit message	Last commit date
Latest commit History 20 Commits
docs		docs
icons		icons
lib		lib
prompts		prompts
public/animations		public/animations
src		src
.gitignore		.gitignore
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
manifest.json		manifest.json
package.json		package.json
postcss.config.js		postcss.config.js
tailwind.config.ts		tailwind.config.ts
tsconfig.json		tsconfig.json
visualizer-sandbox.html		visualizer-sandbox.html
vite.config.ts		vite.config.ts

Folders and files

Latest commit

History

Repository files navigation

Crab-Agent

Table of Contents

Overview

Architecture

Component Map

Background (src/background/index.ts)

Agent Loop (src/core/agent-loop.ts)

LLM Client (src/core/llm-client.ts)

Message Manager (src/core/message-manager.ts)

Tools (src/tools/)

Other core modules

Side Panel (src/sidepanel/)

Agent Loop

Message Flow

execution_event states

LLM Providers

Tool Registry

External Tools

Internal Tools

Memory System

Storage

Entry Structure

memory Tool

System Prompt Injection

Dream Consolidation

Permission Model

Modes

Permission Types

Grant Durations

Domain Safety

Plan Approval

Scheduler

Workflows

Recording

Saving

Playback

Analysis

Tab Groups

Quick Mode

Commands

Configuration

Tech Stack

Build and Install

Prerequisites

Commands

Loading into Chrome

Chrome Permissions

Adding New Features

Adding a new tool

Adding a new LLM provider

Adding a new message type (panel ↔ background)

License

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Background (`src/background/index.ts`)

Agent Loop (`src/core/agent-loop.ts`)

LLM Client (`src/core/llm-client.ts`)

Message Manager (`src/core/message-manager.ts`)

Tools (`src/tools/`)

Side Panel (`src/sidepanel/`)

`execution_event` states

Packages