Introducing Cendor: production plumbing for LLM applications

You shipped an LLM agent. Then production happened: prompts overflow the context window, cost is a black box, every test hits a paid non-deterministic API, and when something goes wrong there’s no record of what the agent saw, did, or refused. Agent frameworks decide what your agent does — they don’t handle these cross-cutting, under-the-call concerns.

Cendor does. It’s one stack with two doors, sharing the flat cendor.* / @cendor/* namespace across Python and TypeScript.

Door one — the libraries

Seven small, composable libraries that plug in beneath LangChain, LlamaIndex, or any SDK:

• contextkit — assemble context to a token budget, with a kept/shrunk/dropped receipt.
• squeeze — reversible, content-aware compression.
• tokenguard — pre-flight budget caps and per-feature/user cost attribution.
• guardrails — a deterministic gate at four stages: block, redact, or flag before the call.
• cassette — record a whole run once, replay it forever offline.
• acttrace — a tamper-evident, offline-verifiable audit chain.

…all on a tiny shared core. You wrap your client once with instrument(); every tool subscribes to one in-process event bus — none monkey-patches your client, none imports another.

pip install cendor-libs        # or: npm i @cendor/libs

Door two — the agent SDK

cendor-sdk is a governed, provider-agnostic agent SDK where budgets, tamper-evident audit, PII redaction, and record/replay testing are the foundation, not plugins — a governed agent in about 10 lines, built on the seven libraries so you can always drop down to them.

pip install "cendor-sdk[openai,anthropic]"    # or: npm i @cendor/sdk

Honest numbers

Every claim is backed by the reproducible benchmark suite — offline, no network, no API keys: uv run python benchmarks/run_all.py.

• 0% token-count error vs the real tokenizer (OpenAI, with tiktoken as a required dependency) — exact, not estimated.
• 97% smaller logs — compressed reversibly, restored byte-for-byte.
• ~15 µs instrument() overhead per call (bus emit + usage extraction + Decimal pricing).
• 1 byte — a single edited byte breaks the audit chain and verify() fails.
• Money is always Decimal / decimal.js, never a float.

acttrace produces evidence to support a compliance case — it is not a compliance guarantee, and not legal advice.

Get started

• Docs: cendor.ai/docs — one page per library, plus the parity matrix.
• Try it in your browser: cendor.ai/try — the real packages running on Pyodide, no server.
• Recipes: the cookbook — every recipe runs offline, no key.

Local-first. No servers. Apache-2.0. Two doors, one honest stack.