Building with LLMs

Why Pixeltable Is Easy to Vibe-Code

Pixeltable’s API is declarative: you say what you want, not how to wire it up. That means LLMs get it right on the first try. Ask your AI tool to “summarize articles with GPT-4o-mini” and you get working code:

import pixeltable as pxt
from pixeltable.functions.openai import chat_completions

t = pxt.create_table('app.articles', {'title': pxt.String, 'body': pxt.String})

t.add_computed_column(response=chat_completions(
    messages=[{'role': 'user', 'content': t.body}], model='gpt-4o-mini'))
t.add_computed_column(summary=t.response.choices[0].message.content)

t.insert([{'title': 'Climate Report', 'body': 'Global temperatures rose 1.2°C ...'}])
t.select(t.title, t.summary).collect()

Ten lines of code, and the result is persistent, versioned, traceable, and incrementally optimized. Every output is stored, every transformation is replayable, and new rows only recompute what changed. The same pattern scales to RAG pipelines, video frame extraction, tool-calling agents, and semantic search.

Set Up Your AI Tool

Pick the setup that matches your editor. These aren’t mutually exclusive; use whichever combination helps.

Agent Skill (recommended, works everywhere)

The Pixeltable Agent Skill teaches AI coding assistants to write correct Pixeltable code on the first attempt. It provides anti-pattern deflection (no LangChain, no pandas-as-store, no for-loops calling AI), correct patterns for 25+ providers, and production recipes for agents, RAG, and multimodal pipelines.

npx skills add pixeltable/pixeltable-skill

Works with any tool that supports the Agent Skills specification: Cursor, Claude Code, Windsurf, Cline, OpenCode, Codex CLI, and 40+ more.

Cursor / Windsurf
Claude Code
Any LLM

The npx skills add command above is the recommended setup. It installs the full skill with anti-pattern deflection, provider coverage, and progressive reference loading.Alternatively, drop our AGENTS.md into your project root for contributor-focused context:

curl -o AGENTS.md https://raw.githubusercontent.com/pixeltable/pixeltable/main/AGENTS.md

The npx skills add command works with Claude Code. You can also install as a Claude Code plugin for auto-updates:

/plugin marketplace add pixeltable/pixeltable-skill
/plugin install pixeltable-skill@pixeltable-skill

The skill loads a concise SKILL.md first (~480 lines), then pulls in reference files on demand only when the task requires them.

Append .md to any docs URL to get a plain-text version optimized for LLMs. Paste it straight into your chat.

Resource	URL
Any docs page as markdown	`https://docs.pixeltable.com/<path>.md`, e.g., this page
Site index for LLMs	llms.txt (standard)
Full site map with metadata	llms-full.txt

MCP Servers

Connect your AI tool to Pixeltable directly via the Model Context Protocol. We ship two servers, or you can build your own using pxt.mcp_udfs().

Docs Search (Hosted)
Developer Server
CLI for agents
Build Your Own

Search the full documentation from Claude Desktop, Cursor, or Windsurf:

https://docs.pixeltable.com/mcp

Exposes a SearchPixeltableDocumentation tool that returns relevant content, code examples, and direct links.

32 tools for creating tables, running queries, managing dependencies, and executing Python, all from your AI editor. Experimental; great for prototyping.

# Install
uv tool install --from git+https://github.com/pixeltable/mcp-server-pixeltable-developer.git mcp-server-pixeltable-developer

# Add to Claude Code
claude mcp add pixeltable mcp-server-pixeltable-developer

See configuration for Cursor, Claude Desktop, and more in the repo README.

The pxt CLI (v0.6.5+) exposes the same catalog with --json on every inspection and query command, plus pxt shell for multi-command sessions. Agents can inspect tables, count rows, and debug computed-column failures without importing Python.See the Working with the Pixeltable CLI cookbook.

Any Pixeltable UDF or query function can be exposed as an MCP tool with a single call:

import pixeltable as pxt

@pxt.udf
def lookup_customer(name: str) -> str:
    """Look up customer info by name."""
    t = pxt.get_table('app.customers')
    return t.where(t.name == name).select(t.info).collect()[0]['info']

tools = pxt.tools(lookup_customer)

pxt.tools() wraps your functions so any MCP-compatible client can call them. See the MCP integration guide for the full setup.

Start Building

Scaffold a full project in one command with pixeltable-new. It generates a working Pixeltable project (serving, backend, or batch) with schema, configuration, and deployment configs already wired up. Ask your AI tool to customize it from there.

uvx pixeltable-new myapp              # declarative serving (default)
uvx pixeltable-new myapp --backend    # full FastAPI + React app
uvx pixeltable-new myapp --template knowledge-base my-kb   # vertical template

Pixeltable Starter Kit

Reference implementations for all three deployment patterns, plus 7 vertical application templates

Next Steps

Quick Start

Install and run your first pipeline in 5 minutes

Computed Columns

The core pattern LLMs generate. Learn how it works

Tool Calling

Build agents with UDFs, queries, and MCP tools

Agents & MCP

Full use case walkthrough for AI agents

Welcome to Pixeltable

Tools

Core Concepts

How-To

Why Pixeltable Is Easy to Vibe-Code

Set Up Your AI Tool

Agent Skill (recommended, works everywhere)

MCP Servers

Start Building

Pixeltable Starter Kit

Next Steps

Quick Start

Computed Columns

Tool Calling

Agents & MCP

​Why Pixeltable Is Easy to Vibe-Code

​Set Up Your AI Tool

​Agent Skill (recommended, works everywhere)

​MCP Servers

​Start Building

Pixeltable Starter Kit

​Next Steps

Quick Start

Computed Columns

Tool Calling

Agents & MCP

Why Pixeltable Is Easy to Vibe-Code

Set Up Your AI Tool

Agent Skill (recommended, works everywhere)

MCP Servers

Start Building

Next Steps