Connect your agent to Grade.
Then ask Claude, ChatGPT or Codex which creators still need paying, what a campaign earned, or what is left in the wallet.
Give this to your agent
Set up Grade: https://usegrade.com/SKILL.mdand let it take it from there.
Your own software can talk to Grade. Everything your team sees, with a key you control.
Your developers get one key from the Developer page in Grade and you decide what it may see and what it may change. Payouts, creators, posts and the wallet are all there to read; sending money needs a key you have allowed to, and a repeated request never pays anyone twice.
Read the docsAsk Claude in plain words. Which creators still need paying, what a campaign earned, what is left in the wallet.
Give your agent the address above and it sets itself up, or paste this configuration into Claude, Cursor or Claude Code yourself. From then on it answers from your Grade. It only reads until you switch changes on, and a payout it prepares is shown to you, recipient, amount and fee, before anything is sent.
Read the docsPut Grade's assistant in your own app. It answers your users and proposes changes; a person confirms.
If you build a product for people who run creator programs, the assistant your team uses in Grade can answer inside yours, with your model and under your name. A change it suggests waits for someone to press confirm, a “yes” typed in the chat is not enough, and nothing it does moves money.
Read the docsModels production serves only certified ones
| Model | x-model | Provider | Status |
|---|---|---|---|
| Claude Sonnet 5.5 | claude-sonnet-5-5 | Anthropic | Certified |
| GPT-6.1 Terra | gpt-6.1-terra | OpenAI | Being certified |
| GPT-5.6 Terra | gpt-5.6-terra | OpenAI | Reserve, tried if GPT-6.1 Terra does not pass |
Limits your provider bills the model
| Limit | Value |
|---|---|
| Questions per end user | 30 in the last hour and 100 a UTC day |
| Questions per application | 1,000 a UTC day, all end users together |
| Commits per application | 60 a minute |
| Creators or links per proposal | 20 |
| One answer | 2 minutes, 6 model calls, about 150,000 input tokens per call, and an estimated US$0.50 of model use |
| Runs at a time | One per end user |
Nothing to install. Claude connects with your Grade sign-in and reads what you can read.
For Claude on the web, or an agent your developers host, Grade runs the connection itself: sign in once in Grade, under Account settings, and your agent reads what you can read for up to 30 days. It can look but not change anything; a change still goes through the assistant and a person's confirm. In beta, for organizations that pay out through Grade.
Read the docsKeep your own copy of the numbers. Payouts, creators and posts, updated as they change.
Finance wants it in the warehouse, the client wants it in their dashboard. Your developers can pull every payout, creator and tracked post into your own database and keep it current, fetching only what changed since last time. Nothing is missed and nothing is counted twice.
Read the docsFeeds every request carries the two Grade headers
| Feed | Scope | Per | Record identity |
|---|---|---|---|
| Campaign media | campaigns:read | one campaign | campaignCreatorMediaId |
| Campaign creators | campaigns:read | one campaign | id (campaign creator record) |
| Contract versions | campaigns:read | one campaign | id (contract history version) |
| Media eligibility scores | campaigns:read | one campaign | id (campaign media record) |
| Payouts | payouts:read | your agency | id (stable payout sync identity) |
| Transactions | payouts:read | your agency | id (transaction) |
| Tracked media | analytics:read | your agency | id (tracked media record) |
What people ask about the API, MCP and agents.
Can Claude pay my creators through Grade?
Not on its own. The assistant API and the reads over MCP offer no action that moves money: a payment an AI starts in Grade is only ever a proposal that a person approves in Grade, with two-factor authentication. Quick Pay and paying payouts stay in Grade.
The MCP server is different, and it is yours to decide. Its write tools are not registered unless you set GRADE_MCP_ENABLE_WRITES to true. When you do, a payout goes trigger, preview, confirm: the client asks with requireConfirmation true and an idempotency key, Grade answers with the recipient, amount, currency and platform fee, and only a second call confirms it. Keep your client’s tool-approval prompts on, so a person reads that preview before the confirm goes out.
How do I connect Grade to Claude, Cursor or another AI client?
Open the JSON configuration your MCP-compatible client uses and add the Grade server: command npx, args -y @usegrade/mcp-server@latest, and your GRADE_APPLICATION_ID and GRADE_API_KEY in its env object. Restart the client and ask it for your Grade wallet balance to check the connection. It needs Node.js 20 or newer, and the same entry works in any client that starts MCP servers over stdio.
A person can also connect Claude without any key: the assistant’s reads are an MCP server at https://mcp.usegrade.com/mcp, and Claude connects to it with that person’s own Grade sign-in, reading as them for up to 30 days. They find it in Grade under Account settings, Connected apps, and disconnect it there. This part is in beta.
Is it safe to give an AI my API key?
Treat the key like a password and it stays one. With the MCP server it lives in the client’s environment configuration, never in a chat message or a tool argument, and the server runs on your machine. With the assistant API, the SDK key, the agent token and your model key all stay on your backend; your interface only ever talks to your backend.
What a key can reach is bounded twice. Its scopes decide which domains it reads or writes, and the MCP server exposes only allowlisted SDK v1 endpoints: it accepts no arbitrary path or host from a tool call. Read tools are on by default and write tools are off until you enable them. Rotating a key keeps its scopes.
Grade MCP server or Zapier and Make for creator payouts?
They do different jobs. Zapier and Make run a fixed recipe when something happens: a trigger, then a sequence of steps you drew in advance. The Grade MCP server gives an AI client tools, so a person can ask in words (“what do we owe on Summer Launch”, “which creators have no payment terms”) and the client picks the reads, and, if you enabled writes, prepares the change for you to approve.
For a schedule that moves data without anyone asking, use the API directly: the same /sdk/v1 endpoints the MCP server calls, and the sync feeds for a local copy. An automation tool that can send HTTP requests with the two Grade headers can use them too.
Which models does the assistant API serve, and who pays for them?
Claude Sonnet 5.5 (x-model claude-sonnet-5-5, Anthropic) is certified. GPT-6.1 Terra (gpt-6.1-terra, OpenAI) is being certified, and GPT-5.6 Terra (gpt-5.6-terra) is the reserve, tried if GPT-6.1 Terra does not pass. Production serves only certified models: Grade runs its evaluation suite through the assistant with a model before it certifies it. During the beta, Grade’s development environment also serves the OpenAI models as candidates.
You bring your own key for the model’s provider in the x-model-key header. It is used for that turn only and never stored, and your provider bills every question to your account. The turn_completed event carries an estimated cost_usd; the real bill is your provider’s.
What are the limits?
Assistant API: 30 questions per end user in the last hour and 100 a UTC day; 1,000 questions per application a UTC day; 60 commits a minute per application; 20 creators or links per proposal; one answer may take 2 minutes, 6 model calls, about 150,000 input tokens per call and an estimated US$0.50 of model use; one run at a time per end user. Over a limit, the stream ends at once with budget_exhausted and says when it opens again.
SDK API: payout currencies USD, CAD, GBP and EUR, amounts positive with at most 2 decimals; payout batches of up to 100 items; campaign creator reads of 100 (v1) or 500 (v2) per page; eligibility score batches of 500 media items; a pending payout lasts 7 days; exports need from and to and stop at 10,000 rows. Sync pages take 1 to 100 IDs and aim at 1 MiB of record data.
How do I get an API key and set its scopes?
In Grade, open Developer, then your application, and create or edit a key. Its scopes are chosen when the key is created and can be changed afterwards. A scope is written domain:action, with read or write as the action, over the domains analytics, campaigns, payouts, approvals, disputes and kyc. Full access includes everything; Custom lets you set each domain to no access, read only, write only, or read and write.
The assistant is a scope of its own, shown as Assistant (agent:read): a key needs it to mint an agent token and to commit. Existing keys do not get it on their own. Scopes are copied into each token, so a change applies from the next one.
Can a key read but not write?
Yes, and that is the default worth starting from. Read and write are separate scopes and write does not imply read, so a domain has four states. A call whose key is missing the scope answers 403 with a message that names it, such as “This API key is missing the required scope: payouts:write”.
The HTTP method is not the rule. Several POST endpoints are pure reads and need only :read, among them POST /payouts/status, POST /payouts/pending and POST /wallet/check-payout-balance, so a read-only key can still check a balance before a payout is approved elsewhere. GET /timezones and GET /settings answer for any valid key.
How do I keep a local copy of payouts in sync?
Use the payouts sync feed, which needs payouts:read and covers your agency. Start without a cursor and an empty local collection; for each page, replace or insert every record in data by its id, delete every entry of removedIds, and commit nextCursor in the same local transaction. While hasMore is true, ask for the next page at once; when it is false, keep the cursor and poll whenever you like.
An unchanged feed returns an empty page, and you save its cursor too. Cursors do not expire by age, so pausing, restarting or rotating the key needs no full reload. Delivery is at least once and records are complete replacements, so applying a page twice is harmless. On an error, retry the last committed cursor and keep your data. Transactions, tracked media, campaign creators, contract versions, eligibility scores and campaign media have feeds of their own.
Is there an OpenAPI file?
For the assistant API, yes: its two SDK routes, POST /sdk/v1/agent/token and POST /sdk/v1/agent/commit, are described in an OpenAPI document, agent/openapi.json, with the production server https://infra.usegrade.com/api/email and a sandbox at https://infra-sandbox.usegrade.com/api.
The rest of the SDK v1 reference is written page by page in the docs, with the request, the response and the scope of every endpoint. The MCP server’s tools follow the same endpoints, so a tool’s page is its endpoint’s page.

Describe your program.
Grade turns it into a program it can track and pay.