Open source · Apache-2.0 · Python + TypeScript

Your AI agents.Under control.

Agents are easy to demo and hard to defend. Cendor is open-source governance your engineers embed in code — small libraries and an SDK beneath any agent framework. Spend is capped before it happens, attacks are blocked before the model sees them, and every action leaves a receipt.

More scenarios — finance, security, board, product: Why Cendor →

Governed session · receipt

support-chat · session #8842 · anonymous visitor · session budget $5.00

  • tokenguard14 product questions answeredexact cost, attributed to this session$1.87
  • tokenguardscripted query flood — cap $5.00 reachedstopped pre-flight — the next call never ran; your app's rule took over: hand off to a humanBLOCKED
  • guardrails1 injection attemptgated before the model saw it$0.00
  • cassettesession recordedreplayable — for the dispute, the postmortem, the test suite
acttraceverify() — chain intact ✓sha256:9f3a1c…e77 · change one byte and this fails
The libraries need no server — this receipt is computed in-process. When you want to watch live, the optional monitor runs on your machine, not ours. This receipt is produced by the libraries your engineers embed. cendor.ai/try runs this exact pattern — every anonymous session is capped. Try it, keyless →
Open source Apache-2.0 — no vendor lock-inLocal-first your data never leaves your stackBenchmark-backed every public number reproduces offlineEvidence packs EU AI Act · ISO 42001 · GDPR · NIST AI RMF
contextkit.pack used 7,200 of an 8,000-token budget · within budgetsqueeze.compress kind="log" · 97% smaller, meaning kepttokenguard.track $1.87 attributed to session #8842tokenguard.projected cap $5.00 reached → BLOCKED, the next call never ranguardrails.gate stage=input · injection attempt → DENIED · $0.00cassette.replay session recorded → replayable · 0% API spendacttrace.verify chain intact ✓ · change one byte and this failscore.instrument one seam under every tool · ~15 µs overhead
Measured, not promised

Numbers a benchmark can reproduce

Every figure on this site is produced by a script you can run offline. These four are the headline set — the method and the caveats are on /benchmarks.

0%
API spend in a replayed CI suite
97%
smaller after squeeze, meaning kept
~15 µs
pre-flight budget overhead
1 byte
of tampering breaks the audit chain
The questions your board will ask

Watch it go wrong. Then watch it handled.

Real product surfaces — a support chat, an internal assistant, a procurement questionnaire, an incident review — each played twice: without Cendor, then with it. About two minutes each.

CFO × CISO · Customer-facing AI — the one every business has right now

Every visitor can spend your money.

Your website has a support chat now — everyone's does. Behind it sits a model with the project's monthly budget; call it $500. Without session budgets, every anonymous visitor is holding that whole wallet: one scripted abuser can drain it in a weekend while your real customers get “try again later.” Cendor caps each session pre-flight — say $5 for anonymous visitors, more once they sign in — attributes every cent to its session, and your app decides what happens at the cap: hand off to a human, or end the session. The attacker hits the wall in minutes and leaves.

→ $500 of exposure becomes a $5 blast radius — cendor.ai/try runs this exact pattern

CISO · DPO · Security

The paste that almost left the building.

An employee pastes a customer sheet into the internal assistant “to clean up the formatting” — names, phone numbers, a card number, headed to an external provider. Cendor redacts or blocks it before it leaves your stack, and the redaction itself goes on the record: your DPO holds proof the control fires.

→ the paste never left — and you can prove it

CEO · Board & Legal

The questionnaire that stalled the quarter.

Your biggest deal hits procurement's AI questionnaire: control spend? prove actions? handle personal data? “We're working on it” slips the quarter. With Cendor the receipts are the answers — enforced budgets, a tamper-evident record, replayable evidence. Stated honestly: evidence to support your answers, never a certification.

→ answered from receipts, not promises

CTO · VP Engineering

The refund nobody approved.

At 2 a.m. your agent refunds the wrong customer. Your logs show what the app did — not what the AI saw when it decided. Cendor's receipt shows saw, did, and refused, tamper-evident, and the recorded run replays the mistake exactly.

→ evidence, not guesswork

More scenarios

The invoice · the injection · the regulator · the coin-flip CI run.

Four more, answered the same way — plus the plain-language “What is Cendor?” for whoever signs off. They all play out on Why Cendor, filtered by your role.

→ every scenario, one place

How it lands in your stack

Your framework stays. Three steps, no rewrite.

01 · WRAP

One line for your engineers

They wrap the AI client once — instrument(client). Overhead measured in microseconds. Nothing else changes.

02 · GOVERN

Rules run before every call

Budgets, gates, and policies check pre-flight — an over-budget or unsafe call is stopped before it costs anything.

03 · PROVE

Receipts for everything

Cost attribution, replayable tests, and a tamper-evident audit trail — ready when finance, security, or a regulator asks.

Engineer? The code lives on the library pages — see how it composes →

Cendor Monitor — OverviewCendor Monitor — Overview (light)+ WATCH · optionalWant to see it all in one screen?One docker run gives you Cendor Monitor — a free, self-hosted journey view that shows every run, its cost per step, and the exact moment a budget or guardrail fired. Standard OpenTelemetry, on your own machine — never a Cendor server. See the monitor →

Govern the call, not the framework.

$pip install cendor-libs

or npm i @cendor/libs · npx @cendor/init

Watch

These normally open in a popup player. With JavaScript off, they play inline below.

The chatbot with an open wallet
The chatbot with an open wallet
The paste that almost left the building
The paste that almost left the building
The questionnaire that stalled the quarter
The questionnaire that stalled the quarter
The refund nobody approved
The refund nobody approved
The seven libraries — video tour
The seven libraries — video tour