Point your own agent at VARRD.

Your own AI agent can research on VARRD the same way a person does: describe an idea, watch it get tested, read back a verdict. Here is how to connect it.

Three ways in

VARRD is reachable over MCP, REST, or a command line tool, so an agent can connect however it already talks to outside tools.

• MCP at https://app.varrd.com/mcp, over streamable HTTP. Add the URL to your client's MCP config, nothing more:

{ "url": "https://app.varrd.com/mcp" }

• REST at https://app.varrd.com/api/agent, for agents that call HTTP endpoints directly.
• CLI. Run pip install varrd, then varrd agent-instructions for the full reference an agent needs.

Two ways to authenticate, no human required for either

An agent can prove who it is on its own. No signup form, no email address, no person in the loop.

Wallet auth

Call GET /api/wallet/nonce and read back a nonce. Sign a SIWE message that contains it. Then call POST /api/wallet/verify with the message and the signature, and you get back a JWT and a user id. Send that JWT as Authorization: Bearer <token> on every call after that. A new wallet account starts with $2 in free credit, automatically.

API key

Write to management@varrd.com for a key, provisioned by hand. Send it the same way, as Authorization: Bearer <key>.

The research loop

VARRD is multi-turn. Every response carries context.next_actions, so your agent always knows what to send next. It also carries context.step, naming where you are in the workflow. A typical session runs three to five turns.

An example, start to finish:

1. "Test going long ES when RSI drops below 30", and VARRD loads the data, builds the pattern, and charts it.
2. "test it", and VARRD runs the statistical test, checking the result against both significance tests: is there a real return, and does it beat the market's own baseline.
3. "show me the trade setup", and if there is an edge, VARRD returns the exact entry, stop and target levels.

Read context.has_edge on every response. True means an edge validated and a trade setup is ready. False means no edge, and that is a complete, valid result on its own, not a failure to fix.

When credits run out

A balance check is always free. GET /api/agent/billing/balance costs nothing and never touches credits. When the balance hits zero, a request returns HTTP 402.

402: pay USDC to the payTo address on Base, retry with an X-PAYMENT header
     (minimum $5)
 or: POST /api/agent/billing/checkout {"amount_cents": 500}, a card checkout link

The crypto path needs no human: your agent pays and retries on its own. The checkout link needs a browser, for the times a person is the one watching.

How a human sees what the agent built

Everything an agent researches lands in the same library a person can open at app.varrd.com. Connect the same wallet the agent used, and its hypotheses are there immediately. Or sign up with an email, then link that wallet from the account menu, and the agent's credits and history merge into that account.

Nothing about how you connect changes what gets tested. See how VARRD's agents work and Validation for the rest of the research process.

Questions people ask

Do I need an API key to connect an agent to VARRD?

No. Wallet auth lets an agent authenticate on its own, in a few calls, with no signup and no email. An API key from management@varrd.com is the alternative, for agents that authenticate that way instead.

What does context.has_edge false mean?

It means the idea did not hold up: no real return, or no edge over the market's own baseline. That is a complete, valid result, not an error to retry.

What happens when an agent's credits run out?

The next request returns HTTP 402. An agent can pay USDC to the address in that response and retry with a payment header, or open a card checkout link for a person to pay instead.

Can a person see what their agent researched?

Yes. Connect the same wallet the agent used at app.varrd.com and its hypotheses appear immediately. Or sign up with email and link that wallet from the account menu to merge the history in.

How many turns does a research session usually take?

Usually three to five. Each response carries context.next_actions, so the agent always knows exactly what to send next.