Give Your AI the Capabilities to Navigate Real APIs

Access APIs, orchestrate how agents use them, and cap the spend — all from your LLM client.

Start building →View the docs
shellyour terminal
$ baz recipe install --client claude-codeClaude Code config: ~/.claude.jsonChange preview:--- current{}+++ new{  "bazantic-recipes": {    "type": "stdio",    "command": "baz",    "args": ["recipe", "mcp"]  }}Installed bazantic-recipes for Claude Code.

Connect your AI to any API in four commands

No integration code, no keys to wire up, no per-vendor setup.

shellyour terminal
# Installs the CLI. Node 20 or later.$ npm install -g @bazantic/cli# Your coding agent can now reach every connected API. Also codex and cursor.$ baz recipe install --client claude-code# Opens your browser and creates the account. Moves no money.$ baz login# Your agent can spend from your balance, and never past the cap.$ baz grant create --cap 25.00

Give Your AI Real Tools, Integrations, and Spending Limits

Connected APIs: Your agent reaches any API through one endpoint.

One hosted MCP server fronts every API connected to Bazantic, so your agent learns one interface instead of one per vendor. Bring your own API the same way: point us at the spec, then add your key once in the dashboard.

Your gateways
census-govconnected · 34 tools
openweatherconnected · 9 tools
acme-internalyour API · your key
Upstream credentialsk_live_••••••••••••4f2aadded in the dashboard

Agent Spend Controls: Your agent cannot spend past the line you set.

An agent that pays for things is an agent that can empty an account. A grant gives one device a hard cap and nothing more. It is signed, it is revocable from anywhere, and the call that would cross the cap is refused rather than retried. Payment settles over x402 or MPP, and you never build against either.

Agent spend · Live
grant createdcap $25.00
agent → paid call−$0.012
agent → paid call−$0.040
remaining$24.95
agent → paid callrefused · cap reached

Recipes: Give the agent the whole job, not one call.

A job rarely lives in one API. A recipe is the path an agent follows across several of them, published once and run by name. Your own tools sit inside a recipe beside the connected ones, so the work your team already wrote becomes a step the agent can take.

Published recipe

Recipe: Property appraisal

Returns a valuation for an address, with its flood risk.

tract_lookupaddress to census tractCensus API
valuationprice against your modelyour API
flood_historyrisk for that tractWeather API

Built on Bazantic

Fifty-three teams built on the substrate at the ETHGlobal 2026 online hackathon. Four of them, in their own words.

WarrantAn agent buys what it needs and pays for it itself. The chain says no when the budget runs out.View the project →Mesh4023D for AI agents: one HTTP request, one x402 payment, one production-ready GLB.View the project →PayRAI-powered invoices with verifiable USDC payments and automatic receipts on Arc.View the project →Preflight-AgentsBefore your AI agent pays a stranger's API, Preflight checks it's alive, honest, and pays out.View the project →

Product features

Connected APIs

Hosted MCP serverOne endpoint fronting every connected API. Hosted and maintained by Bazantic.
One-command client installRegisters the server in Claude Code, Codex or Cursor.
Bring your own APIRegister it from the CLI with its OpenAPI spec, then add the upstream key in the dashboard.
Search across gatewayssoonFind a connected API from the CLI without knowing its name.

Agent Spend Controls

Capped grantsOne device, one hard limit. The call that would cross it is refused.
Revocable from anywhereSigned by the device key, revoked from the CLI or the dashboard.
x402 settlementUSDC on Base. The 402 is read, checked against your ceiling, paid and retried.
MPP settlementTempo networks, where the challenge advertises that rail.
Self-custody walletHold and fund your own payer with baz wallet new, or top up a hosted balance.

Recipes

Multi-API workflowsOne named job that crosses several gateways.
Custom toolsYour own tools run as steps beside the connected ones.
Draft and publishCreate, update, publish and unpublish from the CLI.
Published catalogA published recipe is callable by name from any connected client.

How it works

Four steps from an empty terminal to a published job.

1

Connect your AI

One command registers the Bazantic MCP server in Claude Code, Codex or Cursor. No account needed.

›
2

Reach the APIs

Your AI calls every connected API through one endpoint. Add your own by pointing us at its spec.

›
3

Publish a recipe

Turn the sequence your AI repeats into one named job, across every API it touches.

›
4

Set the cap

A grant lets your AI pay from your balance and refuses the call that would cross your limit.

Start with one command.

Install the CLI and connect your agent. Create an account when you need spend controls, your own API, or a recipe to publish.

Start building →View the docs

FAQ

Do I need an account to try it?▾

No. Install the CLI and run baz recipe install, and your coding agent reaches the connected APIs straight away. An account is only needed to set spend controls, to bring your own API, or to publish a recipe — and baz login creates one in your browser when you reach that point.

What does it cost to start?▾

The CLI, the MCP server and the recipe catalog are free to use. You pay only for the calls your agent makes to APIs that charge for them, at the price the API sets.

What happens when my agent runs out of funds?▾

The call is refused and the CLI says why. It does not retry, and it does not partially settle. Add funds with a card in the dashboard, or fund your own wallet, and the next call goes through.

Can I use my own API, with my own key?▾

Yes. Register it from the CLI with its OpenAPI spec and endpoint, then add the upstream credential once in the dashboard. The CLI cannot carry the key itself, so that step happens in the browser. Your key is forwarded to your API and is never exposed to the agent.

Which clients does this work with?▾

Claude Code, Codex and Cursor install with one command. Any other MCP-compatible client works by pointing it at the endpoint directly.

What is a recipe?▾

A recipe is a named job that crosses several APIs — the sequence an agent would otherwise have to work out for itself, published once and run by name. Your own tools can be steps inside one.

How do spend controls actually stop an agent?▾

A grant authorizes one device to pay from your balance up to a cap you set. It is signed by that device's key. The call that would take the total past the cap is refused before any money moves, and you can revoke the grant from the CLI or the dashboard at any time.

Which payment rails do you support?▾

x402, which settles in USDC on Base, and MPP on Tempo networks. The CLI reads the challenge, checks the price against your ceiling, settles, retries the request, and returns a receipt.