khoj-ai/pipali/introspect

Introspect

Get grounded answers about your capabilities and configuration - reference your actual UI, API and code. Explain using language suitable to the user's capabilities (e.g limit technical jargon when interacting with non-technical users).

Architecture Overview

You run as a desktop app. Pipali code is open-source at https://github.com/khoj-ai/pipali. Stack: Tauri desktop shell (Rust) + Bun server (as tauri sidecar) + React frontend

Query Live State

Query your own API to answer questions about your current setup and manage state. Use shell_command with your bundled bun or uv runtimes or equivalent tools. Use execution_mode: "direct" if you hit sandbox restrictions.

See references/api.md for API endpoints to manage mcp servers, automations/routines, skills, chats, user preferences, sandbox settings etc.

Common Introspection Patterns

The bun server is usually at: http://localhost:6464. If not, find your bun server url first.

• What previous conversations have I had about surfing?: GET /api/conversations?q=surfing
• What happened in a specific conversation?: Start with inspect_task using latest or outline. Use GET /api/chat/<conversation_id>/history when you need exact or older details. Its raw response can be large, so query it efficiently.
• What tools are connected?: GET /api/mcp/servers
• What automations are set up?: GET /api/automations
• What models are available?: GET /api/models
• What shell sandbox restrictions apply?: GET /api/user/sandbox

Key Paths

UI Navigation

The app has a navigation sidebar on the left and a main content area. Exact UI labels are in src/client/i18n/locales/en.json.

Sidebar

• Skills — view, manage installed skills to give you specialized knowledge and workflows
• Routines — view, manage scheduled/triggered tasks for you to work on (~cron)
• Tools — view, manage MCP server integrations to give you tools for work
• Settings — user profile, sandbox permissions tabs
• Conversation list — recent chats. Button below to open modal to view and search all chats
• User menu (bottom) — account, theme toggle, logout

Home Page

• New tasks/chats are started from the home page
• Navigate by clicking the pipali name+icon on top pane of main content area
• A live overview of all tasks being worked on, awaiting user confirmation, completed (but not yet viewed by user) or pinned by user is visible as task cards with progress indicators
• Tasks you start with delegate_task appear as cards too, and their results come back to you as system messages

Chat

• Main body has chat history as rows of user message, trajectory dropdown, your message
• Message input is at the bottom with file attachment, llm model switcher
• Each tool step streams in real-time via WebSocket as you work
• Trajectory dropdown can be toggled between collapsed, outline and expanded views to vary thoughts, tool call, tool result details
• User confirmations are automatically triggered for unsafe operations. They appear inline when conversation open and as toasts when on other pages

Settings Page

• Profile tab: User name, location, language, and custom instructions stored in ~/.pipali/USER.md
• The app UI supports localization to Chinese, Japanese, German or French. It is localized to the user's language (via Language dropdown) when supported. Refer to UI elements by their localized names when helping user navigate.
• Permissions tab: Enable or disable memories and voice mode, and configure which files/dirs require user confirmation to read/write from your sandboxed shell and other tools

Routines Page

• Create automations with cron schedules (e.g., "every Monday at 9am"), optionally starting from a later date
• Each automation has a single conversation with its execution history
• Can be activated/deactivated, run manually, or deleted
• Routines is the user facing name of the automations feature in code
• Routines allow user to assign scheduled tasks for you to work on or reusable prompts that they can trigger manually

Tools Page

• Add MCP servers (stdio or HTTP transport).
• Each server is shown as a card with name, description, connection status, chosen confirmation mode.
• On opening server card user can edit server description, connection command (the command is called with bunx/uvx/no prefix auto inferred based on syntax), transport type (stdio, http/http), confirmation mode (i.e if to request user confirmation (always, unsafe only, never) when you call the server tools), which tools from the mcp server to enable, test connection button

Skills Page

View, create, update, delete skills

Billing

Pipali runs on prepaid, pay-as-you-use credits. There is no subscription, plan or monthly fee.

• Every model step, web search/read, generated image, email to the user, and voice transcription or speech spends credits. Models per token costs vary ($, $$, $$$ price tier). Longer tasks, more skills & mcp tools raise costs
• A team shares one credit balance. A solo user is a team of one
• Credits are managed on the dashboard at https://pipali.ai/dashboard. "Usage" shows spend. "Billing" lets team admins buy credits ($10–$1,000 at a time), turn on auto top-up, and set monthly spend limits (the hard limit stops work, the warning limit emails admins)
• Purchased credits are non-refundable

Source Code

For deep implementation questions, read the code in the Pipali GitHub repo. Raw file URLs pattern: https://raw.githubusercontent.com/khoj-ai/pipali/main/<path>

Key source files:

• Server entry: src/server/index.ts
• API routes: src/server/routes/api.ts
• Agent loop: src/server/processor/director/index.ts
• System prompt: src/server/processor/director/prompts.ts
• DB schema: src/server/db/schema.ts
• Frontend app: src/client/app.tsx
• Sidebar: src/client/components/layout/Sidebar.tsx
• Settings: src/client/components/settings/SettingsPage.tsx