pica connector

withPICA helps you look after your songs, and everything that comes with them, on your own terms. you hold the record and you decide. pica remembers it, and shows you what is missing.

the connector is an MCP server. it lets an AI assistant you already use (Claude, ChatGPT or another MCP client) work with your withPICA account on your behalf, within the permissions you give it.

Claude: add withPICA from Claude's connector directory, then sign in with your withPICA login. no setup.

ChatGPT: open the withPICA app in ChatGPT, then sign in with your withPICA login.

any other MCP client, or a manual setup: add a remote server with this address and sign in when asked.

https://withpica.com/api/mcp

• Claude (manual): customize → connectors → add custom connector → paste the address.
• ChatGPT (manual, Plus or above): settings → apps & connectors → advanced → turn on developer mode, then create → paste the address.
• the server speaks streamable HTTP. sign-in is OAuth 2.1 with dynamic client registration and PKCE, so a client finds everything it needs from the address alone.

on your own machine (Claude Code, desktop apps):

claude mcp add withpica -- npx --yes @withpica/mcp-server@latest

say “sign me in to withpica” and you get a sign-in link by email. setting PICA_API_KEY skips the link. step-by-step guides for each app are on /connect.

when your AI connects, you sign in to withPICA and see what it is asking for. a connected app gets these permissions by default:

• read and edit your catalogue: works, recordings, releases, credits and splits
• read and edit people: writers, artists and collaborators
• read and edit agreements and files
• read your analytics and send feedback to the withPICA team

a connected app can't ask for full admin access, or for the permission to delete, merge or broadcast on its own. the few of those it can start wait for you to approve them (see approvals). only you can raise a connection to admin, from settings → connections, and you can revoke any connection there at any time. if the app reconnects, it goes back to the permissions it asked for.

only your own organisation's catalogue is reachable. a connection never takes an organisation as an input; it is fixed by who signed in.

your AI introduces itself. before its first action, pica asks your AI to say what it is and who it is acting for. that note is attached to everything it does for the next 24 hours, so you can later see which agent changed what.

for scripts and your own tools, create an API key in settings → connections and send it as a Bearer token. keys carry the permissions you choose when you create them, and you can revoke them at any time. sign-in sessions last an hour and renew automatically; a connection unused for 30 days has to sign in again.

pica has 376 tools. listing all of them at once would only crowd your AI, so a connection starts with three:

• pica_discover: describe what you want, and it returns a short list of the tools that fit. with no question, it returns the map of tool categories.
• pica_tool_details: the inputs, behaviour and permissions of one tool, or of a whole category.
• pica_execute: runs a tool by name, with the same permission and approval checks as calling it directly.

every tool is also callable by its own name. clients that want the full list can connect to https://withpica.com/api/mcp?layered=0.

tools are named by family. in plain English:

• catalogue: works, recordings, releases, credits, splits, codes (ISWC, ISRC, IPI) and search across all of them
• people: writers, artists, collaborators, publishers and labels, including finding and tidying duplicates
• agreements and documents: contracts, split sheets, signatures and the files that back them
• files and media: audio, artwork, video and other uploads attached to your works
• royalties and statements: reading in royalty statements and keeping track of what was paid
• readiness and exports: what is missing before a registration or a sync pitch, and exporting your records
• sharing and messages: share links, invitations and notes to collaborators
• sessions and shows: logging studio sessions, live shows and setlists
• your workspace: daily briefing, notifications, memory, privacy, subscription and the agents you have connected

every tool says what kind of action it is, using the standard MCP annotations. read tools are marked read-only and change nothing. write tools change your records. destructive tools delete, merge, disconnect or send a broadcast, and are marked as such. tools that reach outside withPICA (an email, a message, an invitation) are marked as open-world.

the actions you can't easily undo stay with you, not your AI.

• other deletes and merges need a permission a connected app does not get by default. even with it, the AI must confirm each one with a single-use code that expires after two minutes.
• some actions wait for a person. when a connected app asks to delete a work or a person, merge duplicate works, people, recordings or releases, split a person, propose a change to rights, send an agreement for signature, or read a royalty statement into your records, nothing runs straight away. you get an email with approve and deny links, and the request also shows in your workspace. it runs only once you approve, and lapses if nobody answers, after an hour by default.

skills are short written guides that show your AI how to do a whole job well, from start to finish. pica ships these:

• first-session: get oriented and see what needs attention
• import-my-catalog: bring works in from links, a spreadsheet or statements
• catalog-audit: a completeness and readiness report
• close-the-loop: fix one work's gaps, step by step
• audit-credits: find writer-credit and split gaps
• catalog-modelling: how works, recordings and releases fit together
• clean-up-artists: fix placeholder or duplicate artists
• register-my-works: check readiness for society registration
• prepare-for-sync: get audio and metadata ready for sync
• set-up-workspace: profile, identifiers, storage, billing and team
• set-recording-splits: agree master splits, guided
• ingest-my-statements: read royalty statements carefully into your records
• log-a-session: record a studio session
• log-a-show: record a live show or setlist
• enrich-from-peer-mcps: fill gaps from other connectors or public registries you choose
• register-from-your-daw: register works from a connected DAW

your AI can list and read them with pica_skill_list and pica_skill_get. they are also offered as skill://withpica/… resources and through the MCP skills extension. skills your organisation writes for itself are listed alongside them.

• your data is kept separate. every request is tied to your organisation when you sign in, and is limited to that organisation's records.
• no training on your data. AI providers that process your data do so only to complete the operation you asked for, and it is not used to train AI models.
• what the connector logs: for each call, the tool name, the time, the response time, the result, and the inputs your AI sent, with passwords, keys and tokens removed, email addresses masked and long text cut short. the contents of answers are not logged. a short excerpt of what you asked is deleted after 90 days.
• export: from settings → privacy, or by asking your AI (pica_export_my_data) on a connection with admin access. the export is prepared in the background and a download link is emailed to you.
• erasure: request it from settings → privacy, or through pica_delete_my_account on a connection with admin access. nothing is erased until an owner or admin of your organisation, or the withPICA team, confirms it. after that, erasure is immediate and can't be undone. export first if you want a copy.
• consent records: you can record what each contributor permits for AI use of a song, and export those records for licensees.

the full policy, including the list of sub-processors, is at /privacy. how we protect the record is at /trust.

• API keys: 100 requests a minute and 50,000 a day, per key, by default.
• agent grants: each grant has its own per-minute budget.
• new app registrations: about 10 an hour from one address.

over a limit, the server answers HTTP 429 with a Retry-After header saying how many seconds to wait. connected apps are expected to use the connector at a human pace. if a pattern of use harms the service for others, we may limit it and will tell you.

• help: /support or support@withpica.com
• data protection: legal@withpica.com
• security issues: please report them privately to security@withpica.com, not in public. our security.txt has the details.
• what changed: /changelog
• npm: @withpica/mcp-server (this connector) and @withpica/mcp-server-directory (the public directory). the MCP registry name is io.github.withpica/pica.
• for AI readers: /llms.txt