Skeletable MCP Server

Skeletable

Tables for people and agents, where every value records where it came from: a linked page, a person, a file or another app, or a model. Use one for a tracker, a dataset or a list. A new table works this way by default. A table can also require proof, as every table in the public library does: then each value links to the source and exact words that state it, or says where we looked. People read tables on the site, and agents read and build them through an MCP server.

Connect

The server takes MCP over HTTP at /mcp. Reading needs no account. Proposing a change or building a draft needs a token from Settings, sent as a Bearer token. An agent can only write to drafts its person runs, each value still passes every check, and a person publishes.

claude mcp add --transport http skeletable https://skeletable.com/mcp

Datasets

A table whose proof is optional is a dataset. Its values say where they came from: a linked page, a person, a file, or a model. On top of that:

• AI columns have an instruction, like Classify {Pitch} as Horror, Puzzle or Racing. The person's own agent answers it for each row through next_ai_cells and fill_ai_cells, so Skeletable never runs a model itself. Each answer records the model and a fingerprint of its prompt, so a changed row shows its answer is out of date. A value a person typed is never overwritten.
• Versions freeze a table with a lasting link, at /owner/slug/v/2 and, for a public table, /data/owner/slug/v2.csv.
• Column summaries show what each column holds: categories with counts, text lengths, number ranges, and where values came from.
• Data links at /d/<key>.csv load a private table into pandas or Google Sheets. ?columns=a,b picks columns and ?names=0 drops row names.
• Forks copy a public table into your account to extend or relabel, and remember where they came from.

What's here

tables/          The tables, how they're scored, and their check results
packages/tables  The table format and its checks
apps/mcp         The server: the MCP endpoint and the JSON the site reads
apps/web         The site, built with shadcn/ui and served by the server
docker/          The server's image, which Railway builds

Licensing

The source code is licensed under Apache-2.0. Published table data is licensed separately under CC BY 4.0, as shown in the data responses.

Review table changes

Each table has Overview, Proposals, and History tabs, with an Open sheet button for the editor. Signed-in contributors can create a titled proposal with up to 20 cell changes, a reason, and a source and exact quote for each value. Bundles remain open for review, including when submitted by an owner. Existing single-cell owner saves still apply immediately after validation.

Account owners and editors can accept all changes in a bundle or close it. Acceptance is transactional and refuses a bundle if any cell has a newer version than the one recorded at submission. Close a stale proposal and submit an updated one. Accepted values remain in the existing version history. Discussion threads, revision restore, and editing an open proposal are not implemented yet.

Develop

Requirements: Node 22+ and pnpm 10 (corepack enable).

pnpm install
pnpm dev

New tables start as JSON files in tables/, and the server loads them into Postgres when DATABASE_URL is set, or serves them directly when it isn't. It refuses to start if a table fails its checks.

Sign-in needs DATABASE_URL, GOOGLE_CLIENT_ID and GOOGLE_CLIENT_SECRET. PUBLIC_URL sets the site's address, and OWNER_EMAIL says who runs the official account.

Database tests run when TEST_DATABASE_URL is set. They wipe that database.