Anthropic

Claude Code + Pixeltable

Agentic coding in your terminal. Understands your codebase, runs commands, and edits files. Give it a multimodal backend: one app.py declares tables, computed columns, indexes, and HTTP routes.

What is Claude Code?

Claude Code is Anthropic's agentic coding tool for the terminal. It reads your codebase, edits files, and runs commands, so it can install Pixeltable, wire up the Skill and MCP servers, and iterate on app.py until the service answers.

Quick start

Four steps. The agent can run all of them when you hand it the setup playbook.

  1. 01

    Install Claude Code

    Requires Node 18+. Sign in with a Claude account after install.

    bash
    npm install -g @anthropic-ai/claude-code
    claude
  2. 02

    Install Pixeltable

    Python 3.11+ required. `pxt init` marks the project root; schema and service commands refuse an app without one.

    bash
    pip install -U 'pixeltable[serve]'
    pxt init
  3. 03

    Add the Pixeltable Skill

    The Skill teaches the agent table/schema patterns, provider integrations, and anti-patterns. Source: https://github.com/pixeltable/pixeltable-skill

    bash
    npx skills add pixeltable/pixeltable-skill
    npx skills list
  4. 04

    Ship a first app

    Declare the schema, serve the HTTP routes, then experiment in the dashboard. Cloud deploy uses the same app.py: set PIXELTABLE_API_KEY, add `[[pixeltable.database]]` with `name = 'pxt://org:db'`, then `pxt db update`, `pxt schema update`, `pxt service update` in that order.

    bash
    pxt service example --out app.py
    pxt schema update app.py my_app
    pxt service update app.py my_app
    pxt dashboard

MCP configuration

Pixeltable has two MCP servers. Connect the one the job needs.

ServerTransportWhat it can do
pixeltable-webHTTPRead-only docs search and integration listing. No table mutations.
pixeltable-developerstdioInspect schemas, query tables, run a REPL against a local or Cloud catalog.

Connect the WebMCP (docs search, no auth)

One command registers the remote server at user scope:

bash
claude mcp add --transport http pixeltable-web --scope user https://pixeltable.com/mcp

Connect the developer MCP (tables + REPL)

Install the server once per machine (step below), then:

bash
claude mcp add pixeltable-developer mcp-server-pixeltable-developer

Install the developer MCP server

Stdio server that inspects schemas, queries tables, and runs a REPL against a local or Cloud catalog. Repository: https://github.com/pixeltable/mcp-server-pixeltable-developer

bash
uv tool install --from git+https://github.com/pixeltable/mcp-server-pixeltable-developer.git mcp-server-pixeltable-developer
# if uv is missing:
# curl -LsSf https://astral.sh/uv/install.sh | sh

Example prompts

What to ask Claude Code once Pixeltable is set up.

Install Pixeltable and set up a table that embeds every image I insert.

Build a RAG pipeline over the PDFs in ./docs and expose a FastAPI route that answers questions.

Transcribe every video in ./media with Whisper, index the transcript, and serve semantic search over HTTP.

Add a computed column that captions each uploaded image with a vision model.

Scaffold app.py with pxt service example, declare the schema, and start the service.

Take app.py and deploy it to my Pixeltable Cloud database pxt://org:db.

Build a background-remover API: image in, background removed, POST endpoint.

Common workflows

Set up a Pixeltable app

  1. 1.Hand the agent https://pixeltable.com/get-started.md and let it run the playbook
  2. 2.It installs the pxt CLI and the Pixeltable Skill
  3. 3.It scaffolds app.py, declares the schema, starts the service
  4. 4.It verifies the endpoint with a curl request

Build a multimodal pipeline

  1. 1.Describe the pipeline: ingest media, run models, index results
  2. 2.The agent declares the table and computed columns in app.py
  3. 3.It adds an embedding index and an HTTP route that serves results
  4. 4.Insert a row; computed columns run and indexes stay current

Deploy to Pixeltable Cloud

  1. 1.Ask the agent to deploy app.py to Cloud
  2. 2.It pauses for your org slug and an sk_... API key (human step)
  3. 3.It runs pxt db update, pxt schema update, pxt service update in order
  4. 4.It calls the hosted endpoint at https://<org>-<db>.svc.pxt.run to verify

Inspect a catalog over MCP

  1. 1.Connect the pixeltable-developer MCP server
  2. 2.The agent lists tables and views in the catalog
  3. 3.It runs read queries and reports schema and row counts
  4. 4.Keep human confirmation on: the developer server can change local catalog state

Tips for success

Hand the agent https://pixeltable.com/get-started.md; it is the canonical setup playbook.
Point the agent at https://pixeltable.com/llms.txt after setup for the one-file app.py Quick Start.
Keep human confirmation enabled for MCP mutations; the developer server can change catalog state.
Let the agent run `pxt schema diff` before `pxt service update` to preview schema drift.
For Cloud deploys the agent pauses for sign-in and the API key; everything else it can do itself.
Run `claude mcp list` to confirm both servers registered before starting a task.
A CLAUDE.md with a link to get-started.md survives across sessions in the repo.

Troubleshooting

The Pixeltable Skill is not visible after install

Reload or restart the agent. Confirm the install path with `npx skills list` and that the skill targets the current agent.

The agent does not list the MCP server

Validate the config file JSON. pixeltable-web is a remote HTTP server at https://pixeltable.com/mcp; pixeltable-developer is a stdio command, `mcp-server-pixeltable-developer`. Verify the server is installed with `mcp-server-pixeltable-developer --help`.

`pxt service check` fails: modifies catalog while imported

app.py must declare TableModel tables and FastAPIRouter routes, not run pxt.create_table() at import. Scaffold with `pxt service example`.

Cloud deploy fails with 404: UDF not found

Order is required: `pxt db update pxt://org:db`, then `pxt schema update app.py pxt://org:db`, then `pxt service update app.py pxt://org:db`.

Frequently asked questions

Can Claude Code set up Pixeltable on its own?

Yes. Give Claude Code https://pixeltable.com/get-started.md: it installs the pxt CLI, adds the Pixeltable Skill, and connects MCP. It pauses only for Cloud sign-in and the org API key, which are human steps.

Which Pixeltable MCP server should I connect?

pixeltable-web (https://pixeltable.com/mcp) is a remote server for read-only docs search. pixeltable-developer is a stdio server that inspects schemas, queries tables, and runs a REPL against a local or Cloud catalog. Connect the one the job needs.

Can Claude Code deploy to Pixeltable Cloud?

Yes, through the pxt CLI. Claude Code runs `pxt db update`, `pxt schema update`, and `pxt service update` against `pxt://org:db`. Creating the organization and the API key stays with you.

Does the Pixeltable Skill work with Claude Code?

Yes. `npx skills add pixeltable/pixeltable-skill` installs it for the current agent and teaches table/schema patterns, provider integrations, and anti-patterns.

Give Claude Code a multimodal backend

One app.py declares tables, computed columns, indexes, and HTTP routes. Insert a row; the pipeline runs.