Your costs are
logged. Now make
them legible.

The daemon is required. It's the local proxy that captures requests on the developer's machine. Cloud is a hosted home for that data. The daemon stays free and runs entirely on your laptop; Cloud is the sync target.

Metadata only: model, token counts, computed cost, project tag, and timestamp. Prompts and responses never leave your machine unless you flip the separate store_prompt_content opt-in, which is off by default.

Provider dashboards show you what you spent on that provider. Halton Meter shows you what you spent across every provider, attributed to the project, member, or client that incurred it, and reconciles those totals against the provider invoice so you can spot drift.

You get a full CSV export of your data. After cancellation, your data is available to download for 30 days. After that, it is deleted. We do not hold it as leverage.

Anthropic (Claude), OpenAI, Google Gemini, and xAI (Grok) ship first-class adapters today. Those four families are the providers the daemon meters. Anything outside the four metered families passes through unmetered, fail-open. Talk to us about additional providers.

On a schedule (default nightly), Cloud pulls billed totals from the provider's billing API — Anthropic and OpenAI today, with a usage-CSV path for the rest — and diffs them against the daemon's readings. Keys are write-only and KMS-encrypted. Variance under 1% is a tick; over 1% gets flagged for review with a per-day breakdown. Reconciliation is a Business+ feature.

No. The daemon is a transparent local proxy. Your tools keep working exactly as they do now. If the daemon stops, traffic falls through to your providers. Nothing breaks.

We reconcile logged totals against your provider billing statements. In practice the largest variance we have logged is $0.84 on a $279.88 month — about 0.30%. The reconciliation methodology is documented at docs.haltonmeter.com.

Pricing is in USD. Provider invoices reconcile in each provider's billing currency.

From Settings → Billing → Cancel plan. Cancellation takes effect at the end of your current billing period. No cancellation fee. Your data export is available immediately.

Your request path doesn't change. The daemon is a local pass-through — the call goes straight to the provider; no remote gateway hop, no SDK wrapper. The meter reads the traffic as it passes to compute cost. If the daemon stops for any reason, traffic falls through to the provider untouched, so your tools never block on us.

A gateway sits in front of your traffic and routes every call through its own infrastructure — which adds a network hop and makes it a dependency for your requests to work at all. Halton Meter observes locally instead: it watches the traffic your machine already sends and never becomes a required link in the chain. You get the cost data without taking on a new point of failure.

One line. On macOS or Linux, run curl -fsSL https://get.haltonmeter.com/install.sh | sh. On Windows PowerShell, run irm https://get.haltonmeter.com/install.ps1 | iex. It's a fail-open bootstrap that puts the daemon in place. macOS is GA; Linux and Windows are first-class but in public beta. If you prefer the Python-tool workflow, the uvx, uv tool, and pipx instructions still work exactly as before — the one-liner sits alongside them, it doesn't replace them.

Solo: 1 seat, 2 paired machines, 7 days of trace history. Team: up to 10 seats, 3 machines per user, 30 days of traces. Business+ removes the seat cap and adds the governance layer; it is scoped per org. Every tier can read the pricing-rate table its costs are computed from.

Three postures. Record, the default, detects, attributes, and logs every call with no alerts and no interruption. Warn emails the workspace owner and raises an in-app alert while the request path flows untouched. Block stops the call locally before it reaches the provider — fail-safe, no Cloud round-trip, and it still applies if Cloud is unreachable. Hard Block is a Business+ capability, off until a workspace opts in via the allow-hard-block flag, and you can dry-run which recent calls would be blocked before you enable it. Record and Warn are broadly available. Policies are not the same as budgets: budgets are warn-only and never block; only the policy engine can stop a call.