What is the Latten SDK?

An open, local-first toolkit to see — and cut — what your AI costs. It reads usage on your machine and computes cost and waste locally; it reports types and counts, never your prompts, code, or data.

How do I see what my coding agents cost?

Install @latten/glow (npx @latten/glow). It reads your local sessions across Claude Code, Codex, and Gemini, shows one unified balance, prints a quiet per-turn line with tokens and cost (at API rates), and nudges you to take a break and stretch your legs. It runs entirely on your machine, never edits your prompts, and the local tool is free.

How do I instrument the AI agents in my own app?

Install @latten/sdk (npm i @latten/sdk) and wrap each call where your app reaches an LLM, your data, or another service. Latten emits one values-free event per crossing — its cost, reach, and PII exposure as types and counts — so spend and blast radius become visible. It never blocks your app.

Can Codex install Latten for me?

Yes. Add @latten/installer as a Codex MCP server with LATTEN_URL. If no token exists, Codex calls provision once, scans the TypeScript repo, opens a reviewable PR, verifies real crossings, and returns the receipt plus claim URL.

What is @latten/redact?

The internal shared core our tools are built on — not a package you install directly. It holds the cost-and-waste detection, the fix catalog, and the wire-safety boundary that guarantees only types and counts ever leave, never values.

What is @latten/adapter-claude-code?

An internal building block (not separately published): the bridge that turns a Claude Code session transcript into a Latten session. It hashes raw content on-device, so the session carries only hashes, counts, and tokens — never your prompts, code, or output.

Does Latten send my code, prompts, or data anywhere?

No. The local tool is zero-egress — it computes on your machine and needs no account. When you instrument an app, Latten emits types and counts only, never values. We count; we don't copy.

Latten

Install · your unfair advantage

Use all the AI. Lose none of the control.

Go full speed on AI and still know two things, always: what it costs, and that your private data isn't leaking. Everyone else picks one — throttle back, or fly blind. You don't have to. Full throttle, more freedom — not less.

Start free, in one line.

Local · zero-egress · no account. Nothing leaves your machine.

npx @latten/glow

What you see

$ latten balance

  ▌ today across your agents

    claude code   251.7M tok
    codex           3.2M tok
    ──────────────────────────────
    today         254.9M tok

    Instrumentation is live — Latten now sees every crossing. Nothing else to set up.

Every agent you use, on one balance — and a nudge to take a break and stretch your legs. After each turn, one quiet line shows what it cost (at API rates). Types and counts, never your prompts, code, or data.

Using Codex?

Let Codex instrument the app and bring back the receipt.

Add @latten/installer as an MCP server with only LATTEN_URL. If no token exists, Codex calls provision, scans your TypeScript repo, opens a reviewable PR, verifies real crossings, and hands you a claim link.

Install with Codex →

The same advantage, at every scale.

Solo

Free

Ship faster. Don't burn out.

See your usage across every coding agent, keep your AI-life balance, and know when to call it. The whole local tool, free.

npx @latten/glow

Team

$199 flat

Move fast, together.

Your whole team's AI on one view — what it costs and where your data flows — the same control, shared, up to 5 seats. So no one loses the thread.

See it live →

Enterprise

By request

Adopt AI everywhere, safely.

What it all costs, and where your data is exposed, on one live graph — plus the fuller picture as you scale. Full throttle without lowering the security bar.

Start free →

Control is the unfair advantage — at every scale.

Built to take nothing.

Observe-only and local-first — it never blocks your work, and the local tool needs no account.

Types and counts, never values. Dollars are shown at API rates, never as your spend.

We count; we don't copy.

The packages.

The two that matter most lead: your bill, and your data.

@latten/cost — cut your LLM bill: on-device detection of recoverable token spend (redundant calls, stuck loops, oversized and uncached context) with copy-paste fixes. Local, zero-egress.

@latten/redact — see the PII and secrets headed into your LLM calls: on-device detection, with reversible, do-no-harm tokenization, values-free. On npm today.

@latten/glow — the local companion for your coding agents (Claude Code, Codex, Gemini): one usage view across them all and a calm nudge to take a break and stretch your legs. Local, zero-egress.

@latten/sdk — instrument the AI agents inside an app you're building: one values-free event per crossing (cost, reach, PII as types and counts). On npm today. Read the guide →

@latten/adapter-claude-code parses Claude Code sessions on-device — the wire-safety boundary that guarantees only types and counts ever leave, never values.

Technical FAQ · Security · Guides