Connect Claude and other AI assistants
Weekwright speaks the Model Context Protocol, so Claude can read your schedule and draft changes — bounded to one organization, and never more access than whoever connected it.
What this is
Weekwright exposes an MCP endpoint. MCP — the Model Context Protocol — is the open standard AI assistants use to talk to outside systems: Claude's connectors are built on it, so are ChatGPT's apps, and so is the MCP support in editors like Cursor. Point an assistant at the endpoint, sign in once, and it can answer questions about your rota and draft changes to it.
https://weekwright.com/api/mcpUsing a standard rather than building one integration per vendor is the whole reason a small team can offer this at all: the same endpoint serves the Claude apps on desktop, web and mobile, Claude Code, ChatGPT and Cursor — and an assistant that adopts MCP next year works on day one with nothing for us to build. Claude is the client we have verified end to end so far.
The thing worth understanding before you connect anything: this is not a separate API with its own rules. It is the same tool layer the Weekwright dashboard runs on and the same one our own scheduling agent uses. Every call an assistant makes goes through the identical validation, permission check, plan gate, audit write and realtime broadcast as a manager clicking a button. There is no parallel path that can quietly behave differently.
Connecting an assistant
Wherever your assistant keeps its connectors, add a custom one and give it the endpoint URL. In the Claude apps that is Settings → Connectors → Add custom connector; in ChatGPT and Cursor it is the equivalent MCP server field. You will be sent to Weekwright to sign in and approve the connection. The approval screen asks two questions and no others:
- Which organization? An assistant is connected to exactly one. If you manage two sites under separate organizations, the assistant sees the one you picked and has no idea the other exists.
- May it make changes? Off by default. Left off, the assistant can read the schedule and answer questions but cannot alter anything.
That is the whole flow — there is no key to copy, nothing to paste into a config file, and nothing to rotate. If you are wiring up a script or a client that cannot do a sign-in flow, a scoped integration key works on the same endpoint as an Authorization: Bearer header.
What an assistant can actually do
Useful things, on a short leash. Reading: who is on this week, where the gaps are, whether a week would pass your compliance rules before you publish it, who could cover Saturday and why, how hours are distributed across the team, a payroll period. Writing, only if you allowed changes: create a shift, move a shift, post an open shift, add a note to someone's record.
That is 29 operations out of the roughly 137 the product has, and the list is curated by hand rather than generated. Four of them write. Asking an assistant to do anything else does not produce an error you have to interpret — the tool is not offered to it at all, so it does not know the operation exists.
What it never gets
Some things are excluded on purpose and cannot be enabled by ticking a box:
- Publishing a week. Publishing notifies every employee. An assistant drafts; a human presses publish.
- Approving time off or swaps. An approval is an answer to a person, not a data edit.
- Inviting people, changing roles, or touching billing.
- Deleting anything, in bulk or otherwise.
- Creating credentials. A connected app cannot mint API keys or connect further apps — a credential that can make credentials is not a bounded credential.
And underneath all of it: an assistant never has more access than the person who connected it. Connect one as an employee and it sees your own shifts and availability. If you are later demoted, the assistant narrows the same day, automatically — and if you leave the organization, it stops working entirely without anyone having to remember it existed.
Everything it does is on the record
Each change an assistant makes writes an audit entry naming you — the person who connected it — and marking that it arrived through an API rather than the dashboard. A change made by your assistant at 2am is as traceable as one you made yourself, and distinguishable from it.
Changes also broadcast in realtime, so a manager with the schedule open watches the draft appear rather than discovering it later.
Disconnecting
Revoke the connection from Weekwright and the assistant stops immediately — its next call is refused. Revoking is also the way to change which organization it is connected to, since an assistant holds one at a time: disconnect, then connect again and pick the other one.
Common questions
- Which assistants does this work with?
- Anything that speaks MCP: the Claude apps on desktop, web and mobile, Claude Code, ChatGPT, Cursor, and the growing list of clients adopting the standard. You add Weekwright as a custom connector by URL in whichever you use. Claude is the one we have verified end to end so far — we will not claim a client we have not tested, and this answer gets updated as each one is.
- Why one endpoint instead of a Claude app and a ChatGPT app?
- Because MCP is a standard and building per-vendor integrations is how a small team ends up maintaining five half-working ones. A single endpoint serves every assistant that speaks the protocol, and one that adopts it next year works on day one with nothing for us to build.
- Do I need a paid plan?
- No. The endpoint is available on every plan, including the free tier, and is rate limited the same way the rest of the API is.
- Can the assistant see other organizations I belong to?
- No. A connection is bound to one organization at approval time, and a request naming a different one is refused rather than silently redirected. The assistant is never told the others exist.
- Can I connect an assistant for my whole team?
- A connection belongs to the person who approved it and carries their access, so each person connects their own. That is deliberate: a shared assistant acting as nobody in particular would make the audit trail useless.
- What stops an assistant from doing something destructive?
- Three independent limits, all of which must pass: the operation has to be on the curated list at all, it has to be within what you approved (read-only or changes), and it has to be within your own role. Publishing, approvals, invites, billing and deletion are outside the list entirely, so no combination of the other two reaches them.
Related
API and webhooks
One RPC endpoint exposes every operation in the product, authenticated by a scoped org key. Outbound webhooks are signed and retried. Reference for both.
Security and the audit log
How tenants are isolated, what the audit log records, and the recovery paths that exist when somebody deletes the wrong thing.
The AI scheduling agent
What the assistant can actually do, how a week gets drafted, why nothing is written without your say-so, and how to talk to it so it gets the week right first time.