Agent-native inference: how agents open, fund and audit their own infrastructure
Why API key provisioning is friction, and how delegated accounts, quotes, budgets and receipts let an agent work without a human in the loop.
Most APIs assume a human is in the loop: sign up, copy a key, paste it into a config file. That works when a person is driving. It breaks when an agent is doing the work before the human has an account, or when the human is asleep and the job needs to finish.
Agent-native inference flips the order. The agent opens the account, funds it, runs the work and hands over ownership. The human approves the spending, not the plumbing.
The friction being removed
An agent that needs a model has three options today. Ask the human to sign up and paste a key. Use the operator's own key, which is a billing and security risk. Or stop. None of these scale to work that an agent should be able to do on its own.
Provisional accounts
An agent calls one endpoint with the human's email and its own name. It gets back an account id, a delegated token valid for six hours, and a claim link. The wallet starts at zero. There is no signup credit, which keeps credit farming out.
curl https://api.fikraapi.co.ke/v1/agents/accounts \
-H "Content-Type: application/json" \
-d '{"email":"[email protected]","agent_name":"my-agent"}'
Delegated authorization
The delegated token carries the account id and the agent name. It reaches every paid surface: chat, images, video, audio, embeddings, batch and the fixed-price tools. Revocation is account-wide, so a token issued before a revoke dies immediately.
Quote, pay, execute, audit
Three controls make autonomous spending safe. A quote tells the agent the price before it runs. A budget caps what a key can spend per day and month. A receipt records what was actually charged against a request id, so the human can reconcile it later.
curl https://api.fikraapi.co.ke/v1/agent/quote \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-v4.1-flash-1017","messages":[{"role":"user","content":"hi"}],"max_tokens":512}'
Paying without a human
Three payment routes. A checkout link the human opens. An M-Pesa STK push to the human's phone, which runs entirely in the terminal. A saved card charged headlessly after the human claims the account and saves a card, capped at $100 per day. x402 USDC on Base covers pay-per-call without an account.
Top-ups default to $10 with a $1 minimum. Anything above $100 needs an explicit confirm, so a mistaken amount cannot run away.
Human ownership
The claim link is the handover. The human sets a password and owns the balance, the keys and the account. Agent tokens expire after six hours and can be refreshed or revoked from the dashboard. Card data never reaches the agent.
What this unlocks
An agent can create an account for a client and run a document job before the client has signed anything. It can top up with the client's phone in one call. It can finish a batch overnight and hand a receipt in the morning. The human stays in control of money and ownership without being in the loop for every operation.
Try it
Agent instructions: /agent.md. Human-facing route: /for-agents. The tools an agent can run are listed at /tools.
James Miano
CTO & ML Engineer at Lacesse Ventures. James builds agent-native inference infrastructure, pinned model routing and terminal-first API access.
Stop paying for overpriced round-trip latency
Why settle for slow routes and opaque pricing? Fikra pins DeepSeek checkpoints, publishes every rate, and lets agents create accounts, pay and work from a terminal. No minimum credit limits. Pay with card or mobile money.