For developers — notes
One open API with 1,085 operations — the same ones the office, the register, the command line and AI assistants use — with keys, dry runs, webhooks, an agent channel and MCP, documented at docs.almondtill.com.
2. One API, every surface
GraphQL over HTTPS: 1,085 operations — 395 reads and 690 writes — on one schema. The back office, the browser register, the command line and the AI assistants are consumers of it; nothing is screen-only.
Notes
The counts are read from the live schema at the render. The reference is generated from the same schema the API serves, so it cannot drift from it.
3. The reference, generated from the schema
132 record types rostered, each with its operations, its types and its errors on its own pages. The schema itself is served as a file, the errors are catalogued, and there is an index written for AI assistants.
the schema itself · the error catalogue · the index written for AI assistants
Notes
Open a family page: the operations, the types, the errors, an example request and response to copy. The AI index is the whole site in one file an assistant can read.
4. Keys and sessions
An API key is minted by an owner with a scope written as rows — the service, the action, allow or disallow — and it expires on a date the owner chose. The secret shows once; your code exchanges it for a session token and never sends it again. People sign in to sessions; tills run on device sessions.
Let an app or your accountant integrate · the guides
Notes
The authentication guide is the key's whole life: who mints it, what it may do, how it turns into a session, how you rotate it, where it must never go.
5. Your first call and the envelope
Every request and answer carries a small envelope: the version you wrote against and the version that served you, the call's own id, what the call cost. Paging is a cursor grammar; errors come in one grammar with the retries named; the rate limits are documented, not discovered.
Notes
The guides, in order: authentication and key hygiene, your first call, paging and the cursor grammar, versions and the envelope, errors and retries, rate limits, dry runs, webhooks and the signature, the agent channel, search.
6. Dry runs
A write can be rehearsed: every check runs, the write is rehearsed against your current records, and nothing is stored. Ask with one flag beside the version; the answer is what the real call would answer, marked as a rehearsal.
Notes
The dry-run guide names the operations that declare a rehearsal — the creates and the edits of your records, and the item wizard's create — and how to read the answer.
7. Webhooks
Subscribe an https endpoint to the events you pick; every delivery is signed with a secret you hold, arrives at least once, and carries the record's id and revision — never its contents. Read the record back through the same API, so what your receiver stores is always the truth.
Send events to another system (webhooks)
Notes
The two promises and the two duties: at least once and unordered, so de-duplicate on the id and order by revision. The signature is a timed HMAC over the raw body; the guide has the verification.
8. The command line
1,085 operations, from the same registry — the command line is a consumer of the API like every other surface, so a script can do what the office does, and nothing is screen-only.
Notes
Show a list, a get and a dry run from the terminal. The command line is where the wizard drafts and the fill conversation came first.
9. AI assistants over MCP
17 tools and 6 resources and resource templates, read-first, with only the permissions the merchant assigns. The assistant signs in over OAuth as a person, or rides an API key; the merchant connects it with one press in the office, or one command in a coding agent.
Notes
Both counts are read from the MCP registries at the render. The handbook and the help pages are resources the assistant can read; every call is tracked and capped by the merchant's wallet.
10. The agent channel
A buyer's AI agent can find your shop, search the catalog and complete a checkout over an open standard. Your side is a set of ordinary operations of this API: create the channel, publish it, pause and resume it, retire it.
Let AI shopping agents buy from you
Notes
The agent channel guide is a pointer: the controls are in the reference; what the agent platforms call is the platform's own channel, secured by the channel's signing keys.
11. Search, feeds and exports
One word finds the records that carry it, every kind at once, only the kinds your key may read. A question becomes a plan you can read, run while you wait or as a job. Scheduled data feeds carry your records to a BI tool; the export is machine-readable and whole.
Schedule a data feed for your BI tool · the use cases
Notes
The search guide: the typeahead, the exact lane for codes, the question lane for anything longer. The feed is the use case; the export is the covenant's promise.
12. What it costs to build on
The kernels meter: 100 million kernels are included inside the $15 a month the merchant already pays, and every call is metered and shown to them in their own ledger — the envelope on every answer says what the call cost.
Notes
The merchant's fee carries the usage, and the merchant sees it. The sample months stand on the pricing page — do not quote them from memory.
13. Next steps
Read the reference, mint a key with an owner, make the first call. Then apply with the merchant — no card, no fee.
the developer documentation · Start your evaluation
Notes
The first call guide has one request to copy. The application is the merchant's; the developer's door is the documentation.
Where these numbers come from
Every figure on these slides is read from the system when the page is rendered — the registries, the rate card, the example shop’s own records — or, for the other side, from dated public list prices. Nothing is typed.
| Figure | Value | Read from |
|---|---|---|
| the monthly fee for the whole business | $15 | the rate card — the base fee and the included kernels, derived at the render |
| the kernels included every month | 100 million kernels | the rate card — the base fee and the included kernels, derived at the render |
| record types the specification rosters | 132 | the construct catalogue — counted at the render |
| resources and resource templates an AI assistant can read over MCP | 6 | the MCP resource registry — the fixed resources and the templates, counted at the render |
| tools an AI assistant reaches over MCP | 17 | the MCP tool registry — the meta-tools and the typed tools derived from the live schema, counted at the render |
| writes among the operations | 690 | the live schema the API serves — counted at the render |
| operations on the one open API | 1,085 | the live schema the API serves — counted at the render |
| reads among the operations | 395 | the live schema the API serves — counted at the render |
The doors: the developer documentation (the developer documentation site) · the error catalogue (the developer documentation site) · the guides (the developer documentation site) · the index written for AI assistants (the developer documentation site) · the API reference (the developer documentation site) · the schema itself (the developer documentation site) · the use cases (the developer documentation site) · the pricing page (the public site) · Start your evaluation (the public site) · Let an app or your accountant integrate (the developer documentation site) · Connect your AI assistant (the developer documentation site) · Schedule a data feed for your BI tool (the developer documentation site) · Let AI shopping agents buy from you (the developer documentation site) · Send events to another system (webhooks) (the developer documentation site).