blend0.

How it works

Every call goes through one place.

Your agent sends its outbound calls to Blend0 instead of calling services directly. Blend0 checks each one against your policy, and when a service asks to be paid, decides before the agent's wallet signs.

01

The flow

  1. Check the host

    Deny rules win, then allow rules, then your default. Blend0 also checks the address the name resolves to. A blocked host is never contacted.

  2. Forward the call

    Allowed calls go through. Free answers stream back to the agent byte for byte.

  3. Decide on payment

    If the service asks to be paid, Blend0 checks price, budget and trust, and answers allow, review or deny.

  1. Agent signs

    On allow, the agent's own wallet signs. Blend0 never sees the key.

  2. Paid retry

    The agent retries with a single-use approval. Blend0 checks it and forwards the payment.

  3. Record the outcome

    Only payments that got a working answer count as spend. Services that fail lose trust.

02

Payment rules

Every decision starts at allow. Each rule can only make it stricter, and each one that fires adds a reason.

RuleResult
Payment network or currency not on your listdeny
Price above the per-request capdeny
Price above what's left of today's budgetdeny
No reputation for this service yetreview
Trust score below your minimumdeny
Price above the auto-approve thresholdreview
None of the aboveallow

Every result is a setting you can change. These are the defaults.

03

Approvals

Each payment decision creates an approval. It works once, for one service and one agent, for five minutes by default. Approvals are saved to disk, so a restart doesn't drop them.

A reviewed payment waits until a person approves or denies it, in the admin console or from the command line. Blend0 can post it to a webhook, such as Slack or Discord, so nobody has to watch.

blend0 pending
blend0 approve 4b0d6f2e

Blend0 signs each approval: who gets paid, how much, on which network, and until when. Wallet code can check that signature with Blend0's public key before it signs, so an agent can't forge or stretch an approval. Blend0 signs its decision, never the payment.

Approved payments hold their amount against the daily budget until they're used or expire, so many agents asking at once can't overshoot it together.

04

Trust

Each paid service has a score from 0 to 1. It starts from a score you set, or from neutral, and is blended with Blend0's own record of what that service returned after being paid. The more payments it has seen, the more its own record counts.

With nothing known about a service, Blend0 asks a person. The record stays on your machine.

05

Agents and budgets

With no setup, every caller is one agent sharing one budget. Give each agent its own key and limits with one command:

blend0 agent add research --budget 1 --max 0.05

Once any agent exists, every call needs a key, and unknown keys are refused. Each agent spends from its own daily budget and the team's. An agent's settings can only be stricter than yours: its deny rules add to the global list, and each limit is the lower of the two.

06

Admin console

Blend0 serves an admin console, on the gateway's port or on its own. Approve payments, search the activity log (and export it, on Team), add agents, edit and test policy, restore earlier versions and manage users. Changes apply at once, with no restart.

RoleCan
ViewerSee budgets, approvals, activity, agents and policy
ApproverAlso approve or deny payments
AdminEverything: agents, policy and users

Sign in with a password, or, on the Enterprise plan, with single sign-on through OpenID Connect: Okta, Microsoft Entra ID, Google Workspace, Auth0 or Keycloak. Groups map to roles, and anyone without a mapped group is refused.

07

Changing the rules

Test first. Before you save a change to network rules, payments or trust, Blend0 can replay the last 30 days of recorded traffic with it: which requests would have been blocked or let through, which payments paid, held or refused, and whether it overturns something a person decided. Nothing is saved until you say so.

Undo anything. Every change, from the console, the command line or the advisor, keeps the version it replaced, with who changed it. Restore any version in one click. Agent keys never roll back.

Suggestions (Team and Enterprise). The advisor reads your team's own record: payments people approved and refused, and paid services that delivered or failed. It suggests rules from that, never from how often an agent asks, so an agent can't talk its way into trust. Each suggestion is replayed on your traffic before you see it; any that would let through something a person refused are dropped. Try one in a shadow test first, which changes nothing, or apply it.

Automatic tightening (opt-in). Turn it on and suggestions that only make the policy stricter apply themselves, up to a number a day. Each is watched, and if it blocks more than you allow, it undoes itself. Anything less strict always waits for a person.

08

Lockdown

Blend0 only protects calls that go through it. To make it the only way out, run the agent on a network with no route to the internet, and Blend0 beside it. The repository has a Docker recipe that does this.

For tools that can't change how they make requests, turn on proxy mode and point HTTPS_PROXY at Blend0. It applies network rules to that traffic too. It sees hosts, not what's inside HTTPS, so payments must still go through Blend0's fetch. Payment headers sent through the proxy are refused.

09

Three ways in

Claude and other MCP clients

One command. Blend0 starts by itself and gives the agent three tools: fetch, payment status and budget.

claude mcp add -s user blend0 -- npx -y blend0 mcp

Your agent's code

A drop-in fetch. Allowed calls return the service's response. Blocked calls and refused payments throw with Blend0's reasons. Payments waiting on a person are polled for you, and approved ones are signed by your wallet code.

const fetch = createBlend0Fetch({ signPayment });
const res = await fetch("https://api.example-data.com/v1/quote");

Any language

One endpoint. Send the call, read the decision.

POST /v1/fetch
{ "url": "https://api.example-data.com/v1/quote" }

← 402
{
  "blend0": {
    "decision": "allow",
    "approval": "4b0d6f2e-…",
    "reasons": ["trust score 0.92", "approved by policy"],
    "option": { "network": "base", "asset": "USDC", "amountUsd": 0.001 },
    "budgetRemainingUsd": 4.82
  }
}

Every request, decision, approval and payment is written to an append-only audit log, with its reasons. So is every sign-in and admin change, with who made it.

Next: security and limits