Connect an agent with MCP

An agent builds and runs your Product through the MCP server at https://hyperscale0.ai/v1/mcp. It works freely in sandbox with test money. Live waits for you.

Connect

You use	Connect with
Claude, Codex, Gemini or Cursor	The URL. The client opens a browser: sign in and choose Allow.
A coding agent in a terminal	`hyperscale agent setup`, which sends the agent's own token.
A service that operates a Product	A Product API key as `Authorization: Bearer <key>`.

claude mcp add --transport http hyperscale https://hyperscale0.ai/v1/mcp
codex mcp add hyperscale --url https://hyperscale0.ai/v1/mcp
gemini mcp add hyperscale https://hyperscale0.ai/v1/mcp --transport http

In Cursor, open Settings > MCP and add an HTTP server named hyperscale with that URL.

A coding agent gets a sandbox-only login in one command:

hyperscale agent setup claude-code
export HYPERSCALE_TOKEN="$(hyperscale agent token claude-code)"

The MCP entry it writes sends Authorization: Bearer ${HYPERSCALE_TOKEN} and never holds the token. Sign in covers agent names, Product access and the Codex and Cursor entries.

The server also takes a personal access token, hyperscale_pat_.... You make one on the account page at https://console.hyperscale0.ai/account/tokens and revoke it there at any time.

Ask

Tell the agent what you sell and who pays whom:

I rent desks by the month. A member pays a SAR 500 deposit and SAR 1,200 on the 1st of each month, with a SAR 50 fee after the 5th. Build it in the sandbox and run one member through three months.

What the agent calls

Step	Tools
Find its company and Products	`agent_member_current`
Build the Product	`composer_check`, `composer_plan`, `product_create`
Add customers and records	`customer_create`, `account_create`, `product_objects_create`
Run an action, such as pay	`product_objects_actions`, then `product_objects_execute`
Fund a test account	`sandbox_account_fund`
Move the sandbox clock	`sandbox_clock_advance`
Check a balance	`account_balance_retrieve`
See the whole company	`company_graph_retrieve`
Snapshot and compare the books	`company_snapshot_create`, `company_snapshot_list`, `company_snapshot_diff`
Import a bank statement	`bank_statement_import`
Clear reconciliation exceptions	`reconciliation_exception_list`, `reconciliation_exception_resolve`
Read what agents did	`agent_activity_list`

Snapshot create, statement import and exception resolve reach every Product, so they run only on a grant from hyperscale agent setup --all-products. A grant on some Products gets 403 for them and never lists them.

tools/list shows 28 everyday tools. The agent finds the rest with tool_list and runs them with tool_call, or connects to https://hyperscale0.ai/v1/mcp?toolset=all to list every one. /.well-known/mcp.json names them all.

Rules

Amounts are strings in minor units: SAR 20.00 is "2000".
Every change carries an idempotencyKey. The same key retries a change and never repeats it.
Every change carries an agentReason: one sentence of 8 to 280 characters on why. You read the reasons in hyperscale agent activity and on any approval.
In sandbox, money tools and the sandbox clock run without a confirmation. A tool that can't be undone, such as product_close, still asks you first.
Any change can run with rehearse: true first. It answers the receipt and ledger postings the real call would write, writes nothing and skips the "ask the founder first" check. A checkpoint tool still needs the same confirmation hash it would run with.
Live waits for you to approve each action in a browser, reads included.
Product API keys come from the CLI, never through MCP: hyperscale key create --product <productId>.

When it fails

Every refusal has one plain reason, a nextStep naming the tool or argument to fix, and a traceId to quote to support.

Refusal	Meaning	Next
`agent_approval_required`	A live action waits for you	Approve the link within 24 hours. The agent retries.
`agent_cap_exceeded`	A money call is over the agent's caps	Run `cliCommand` within 24 hours, or raise the cap.
`agent_reason_required`	A change came without an `agentReason`	The agent sends the call again with one sentence on why.
`agent_token_replaced`	`agent setup --rotate` ran	Run the `export` in `nextStep`, then restart the agent.
`agent_token_expired`	The grant ended: 30 days, or 7 days unused	Run `hyperscale agent setup` again.
`-32602`	An argument is wrong, such as `"20.00"`	The message names the argument. The agent fixes it.