AI AGENT AUTHORITY

Give AI agents authority. Not a blank check.

Fidacy defines exactly what an agent may do, decides whether each consequential action is inside that authority, and creates proof anyone can verify later.

Mandate. Decision. Evidence. For money, the decision can also stop the transaction before it moves.

See a live decision → Connect your agent →
hold the cube. every verdict joins the chain
{{ m }}
{{ m }}

The system that acts cannot be the only judge.

action.executed  crm:customer-4419  credit USD 100.00
record.updated    "was allowed" · written after the fact
audit.log · produced by the system that acted
// THE SYSTEM ACTS AND ATTESTS TO ITSELF
✕The party taking the action produces the record of what was allowed
✕The record can be regenerated after the fact
✕Disputes resolve to the vendor's own logs
DENYbefore execution · ~8ms
Ed25519 · verifiable offline, by anyone
chain: a4f9…e21c → checkpoint → Bitcoin
// FIDACY ISSUES THE VERDICT BEFORE THE ACTION
✓Ed25519 verdict bound to the agent, mandate and exact action
✓Verifiable offline by any third party, without calling Fidacy
✓Hash-chained and anchored to Bitcoin via public checkpoints
✓Custody never leaves the customer

This is not a feature. It is a separation of powers.

// THE REAL BLOCKER ISN'T THE MODEL

What stops agents at scale isn't intelligence. It's liability.

Every company wants agents doing real work. Almost none lets them, because when the agent gets it wrong, a person inside that company owns the loss, the fine, and the explanation.

$ npx -y @fidacy/mcp
✓ authority guard active · default deny
THE BUILDER

"If my agent changes the wrong record, sends the wrong message or pays the wrong person, that's on me. So I review everything by hand, which kills the point of having an agent."

A mandate gives the agent narrow authority. Fidacy decides against it before the action and records the result. For payments, the firewall enforces the same boundary at the money rail.

grant g_7f2aSIGNED ✓
payee  supplier:acme
amount USD 100.00 · invoice #4419
THE CFO

"I can't approve agents that spend money when I can't prove what was authorized. When it goes wrong, the loss lands on my desk."

Every consequential action carries a signed decision bound to the agent and mandate. For payments, an allow also creates a one-time grant bound to the payee, amount and invoice. No valid grant, no settlement.

₿
THE COMPLIANCE OFFICER

"Competitors are scaling agents. We can't, because I have nothing to hand an auditor except our own logs."

Every decision lands in a tamper-evident audit anchored to Bitcoin. An auditor, a regulator or a counterparty verifies it against public keys, without trusting you or Fidacy. Evidence, not logs.

Fidacy exists to take that question off your desk, so the agent can actually work.

Start free →
The evidence layer that makes agents insurable→ Agents at scale, without owning the liability→ Read the open Grant spec→
// DON'T TAKE OUR WORD, RUN IT

Send an agent action. Get a signed verdict. Verify it in your own browser.

approve / review / deny, and a cryptographically signed verdict (EdDSA) you check against our public JWKS with an external library. The signature is valid whether or not you trust Fidacy. That's the whole promise, in your hands, in milliseconds.

Run a live verdict → See the production console →
Live in production. EdDSA-signed. Audit anchored to Bitcoin.
~8ms server-side (warm path). Verifiable by anyone.
$ node quickstart.mjs # fky_test_…
# 1) assess: engine scores, signs
decision : deny (anonymous agent)
score    : 50
# 2) verify the signature YOURSELF
signature valid : true ✓ ← public JWKS
signed == resp  : true ← untampered
# register the agent, same engine:
decision : approve (in-mandate)
signature valid : true ✓
# identity + mandate flip the verdict.▋
ENGINE INTEGRITY · AUDITABLE
{{ s.value }} {{ s.label }}
every figure is counted by the engine and public. re-fetch them yourself: api.fidacy.com/v1/pulse · the full public record, anchored to Bitcoin →
// COURTS, REGULATORS AND REAL LOSSES, ON THE RECORD

The liability isn't a thesis anymore. It's on the record.

Courts and regulators on three continents converged on the same holding: your agent's action is your action. Every row links to the primary source.

{{ r.case }} +

{{ r.body }}

{{ r.link }} →
every decision hash-chained · Merkle-anchored to Bitcoin
// AGENTS ALREADY ACT IN THE REAL WORLD

Microsoft's security team now tells enterprises to treat a popular agent framework as "untrusted code execution with persistent credentials". Here is what that looked like in the first half of 2026:

US$440K sent by one AI agent in a single wrong Solana transfer: a decimal parsing error, live on X Cointelegraph · Feb 2026 →

"In this new payments paradigm, trust becomes the product."

Mastercard, announcing its agent-payments standard · Mar 2026 → 1,184 malicious skills planted in the ClawHub marketplace in one campaign; US$2.3M in crypto stolen Antiy CERT · Feb 2026 → 40,214 OpenClaw agent instances found exposed on the public internet, most of them vulnerable SecurityScorecard · Feb 2026 → 84% of 1,440 fraud-prevention leaders across 25 countries call AI agents the industry's most exploitable vulnerability of the next year BioCatch · Jun 2026 →
Become a design partner → See the attacks it blocks, counted live →
// WORKS WITH THE AGENTS YOU RUN

Drop Fidacy into the agents you already run.

Set a bounded mandate at the action boundary, then connect the runtime you already use. Fidacy records local agent activity and adds a signed decision wherever authority must be checked.

{{ a.name }} {{ a.pkg }} {{ a.body }} Read the integration →

Running something else? The universal connector wires Fidacy into any agent runtime with one drop-in guard. See the universal connector →

Your first signed, verifiable verdict. This afternoon.

Create a mandate, connect an agent, and see the signed decision and evidence trail for yourself. The firewall remains available where money must be enforced.

Open the console → Run a live verdict →
Live on UCP · AP2 · A2A.
IF FIDACY IS DOWN

A money-moving action is denied. The firewall fails closed in every adapter and in the engine itself.

WHAT IT COSTS

Local observer and firewall free. Hosted signed verdicts are metered from $0.04 each. Full pricing →

WHO SIGNS

Built by Lucas de Lima and team. Meet the team →