blocknet.

Agent Gateway · open source

Security and spending control for AI agents.

Agents now call APIs, use tools and pay for services on their own. Blocknet decides what an agent may reach and what it may pay for, within budget and only for services it has reason to trust. Then it keeps those rules right, tested on your traffic before anything changes.

Agent requests a paid API
GET api.example-data.com/v1/quote402
Service reputation
Trusted · 0.92
Requested payment
$0.001 USDC
Budget remaining today
$4.82
Network policy
Allowed
Approved by policysingle use · expires in 5 min
Example values

01

Get started

Safe with no setup. Pick the way in that fits.

  • Claude and MCP clients

    One command, no code

    Add Blocknet as an MCP server. Your agent gets a fetch tool that goes through Blocknet, plus payment status and budget.

  • Your agent's code

    A drop-in fetch

    Swap your fetch for Blocknet's. Allowed calls work as before; blocked ones and refused payments throw with the reason.

  • Any language

    One HTTP endpoint

    Send calls to the gateway and read the decision. Works from anything that can make a web request.

  • Out of the box

    Safe defaults

    Cloud metadata blocked, even behind another name. $0.10 cap per request, $5 a day, and a person asked before paying anyone new.

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

Free and open source under Apache 2.0, on npm, source included. Node 22, no dependencies.

02

The problem

An agent will call whatever it's told to, and pay any service that asks. Most teams can't answer two questions about their agents: what did it reach, and what did it spend?

  • 01

    It reaches where it shouldn't

    A prompt-injected agent can send data anywhere, or read cloud credentials from the metadata address.

  • 02

    It spends without limits

    One runaway loop can drain a wallet a cent at a time, and payments on-chain can't be refunded.

  • 03

    It pays for nothing

    Services take payment and return errors. Nothing records which ones actually deliver.

  • 04

    The rules drift

    Rules that are too loose let risk through. Rules that are too tight get approved by reflex until nobody reads them.

03

One control point

Every outbound call goes through Blocknet. Each layer answers one question, and the strictest answer wins.

  • Network policy

    May this agent talk to this host at all, by name and by the address it resolves to?

  • Payment policy

    Is the price within the cap, this agent's budget and the team's, and the allowed networks and currencies?

  • Trust

    Do we have reason to believe this service delivers?

  • Human approval

    Is this unusual enough that a person should decide? They're told by webhook and answer in the admin console, with roles (and single sign-on on Enterprise), or from the command line.

  • Audit

    What happened, and why was it allowed?

  • Improvement

    Are the rules still right? See below.

How a request moves through Blocknet

04

Rules that improve, safely

Blocking is the floor. Blocknet also keeps its rules right: it proposes changes from what your team actually decided, tests each one on your real traffic, and makes every change easy to take back.

  1. Learn from your team

    Which payments people approved and refused, and which paid services delivered. Never how often an agent asks: an agent can't earn trust by repetition.

  2. Propose

    Block a service people keep refusing or one that takes money and fails. Trust one that always delivers. Cut a budget far above real spending.

  3. Test on your traffic

    Each change is replayed on the last 30 days first: what would have been blocked, paid or held. Anything that would undo a person's refusal is dropped.

  1. Apply

    A person applies it. Or, if you turn it on, stricter changes apply themselves. Anything less strict always waits for a person.

  2. Watch

    Each automatic change is watched. If it blocks more than you allow, it undoes itself and isn't tried again.

  3. Undo anything

    Every change, by anyone, is kept as a version. Restore any one in a click.

Testing changes and undoing them are free on every plan. Suggestions and automatic tightening come with Team and Enterprise. How changes stay safe

05

Principles

  • Deny is cheaper than refund

    Stop a bad payment before it's signed. There's no undoing it after.

  • Non-custodial

    Blocknet never holds keys, signs payments or moves funds. The agent's wallet stays the agent's.

  • Every decision explains itself

    Plain-language reasons the agent can act on and a person can audit.

  • Honest about limits

    We say plainly what Blocknet doesn't enforce yet. Read the limits.

  • Open source

    The gateway is open source under Apache 2.0, and the npm package includes its full source, so your security team can read every rule it applies.

  • Protection is never paid

    Every plan blocks, budgets and approves the same way. Paid plans add what teams need around it.

06

Plans

Free to run. Pay when your team relies on it.

  • Free

    Open source, for anyone.

    • The full firewall: rules, payments, budgets, agents, approvals, audit log
    • Test any rule change on your traffic; undo any change
    • Admin console for up to 3 people
  • Most teams

    Team

    For teams running agents that spend money.

    • Suggested rules from your team's decisions, and automatic tightening
    • Up to 25 people; audit export and metrics
    • Email support
  • Enterprise

    For companies with an identity provider and a security review.

    • Single sign-on with groups mapped to roles
    • Unlimited people
    • Help with rollout; a support agreement
Compare plans

Give your agents a budget, not a blank check.

Blocknet is free and open source. Start in five minutes, or write to us about a pilot for your team.

Get it on npm