Why Cendor
Cendor is open-source governance your engineers embed in code — seven small libraries and a governed agent SDK beneath any agent framework. Spend capped before it happens. Attacks blocked before the model sees them. A receipt behind every action.
Production plumbing for LLM applications — the unglamorous layer beneath your agent framework that controls cost, gates unsafe inputs and outputs, makes agents testable, and keeps an audit trail. Python and TypeScript. Free, open source, Apache-2.0.
One problem each: pack the prompt honestly (contextkit), shrink history reversibly (squeeze), cap spend pre-flight (tokenguard), gate attacks (guardrails), record & replay runs (cassette), keep tamper-evident receipts (acttrace), and give every tool one seam (core). Use one, or all seven — under the framework you already have.
It's the second door: start governed instead of retrofitting. Your team builds the agent on cendor-sdk and every run is budgeted, gated, audited, and replayable by construction — a working governed agent in about 10 lines, built from the same seven libraries.
It removes the four blockers between an AI demo and an AI product: unforecastable spend, injection risk, no audit evidence, untestable behavior. Your feature ships with the answers finance, security, and compliance will ask for — built in, not bolted on.
Both doors are the same seven libraries underneath — starting through either never locks you out of the other.
Door one · You already have an agent
Your team keeps its framework and its code. Engineers add small libraries underneath — no rewrite, no migration, no new platform to run.
Door two · You're starting fresh
Don't retrofit governance later — begin with it as the foundation. cendor-sdk is a provider-agnostic agent loop where every run is budgeted, gated, audited, and replayable by construction, and it grows with you: memory, retrieval, multi-agent, ten provider paths.
Which door? Have an agent → libraries. Building one → SDK. Either way your engineers decide in an afternoon — both install with one command.
And a window · watch it allTwo doors — and a window.One wire, two ways to watch it. Cendor emits standard OpenTelemetry — export it to your own backend (Azure Monitor, CloudWatch, Datadog, any OTLP), or open Cendor Monitor, a free, self-hosted journey view, to see your governed agents in one screen: every prompt, token, dollar, and the exact step where a budget or guardrail acted. On your own machine — never a Cendor server. See the monitor →Real product surfaces, played twice: without Cendor, then with it. Filter by your role. Every card has its own link — send it to the person who needs it.
Honest by design. Cendor produces evidence to support EU AI Act, ISO 42001, GDPR, and NIST AI RMF processes — never a compliance guarantee, never legal advice. Costs are labeled reported vs estimated, in exact decimals. Cendor is an open-source project (Apache-2.0), local-first, with no server and no data collection — your prompts and spend never leave your stack.
Created by Raghav Mishra (PowerAI Labs). See About and Privacy.
These normally open in a popup player. With JavaScript off, they play inline below.