Your developers are burning tokens on noise.

Numbers and a tool name — that's the entire schema. Each metric record contains a timestamp, the first word of the command (max 32 chars, no spaces allowed, so arguments can't leak), a category, which AI agent was in use, raw/filtered token counts, filter timing, and an optional salted project hash. Never code, arguments, paths, output, or environment variables. The allowlist is enforced by an automated test on the agent and a strict schema on the server — an unknown field is rejected, not silently stripped.

No. Filtering adds less than 10ms per command, and metric reporting happens asynchronously in a detached process — it is never in the command path. If the server is unreachable, commands are unaffected.

Claude Code and GitHub Copilot are first-class citizens side by side, along with Cursor and Gemini CLI — in the agent hooks, the dashboard, and the reports. ltk init also supports Codex CLI, OpenCode, and other agents. We cover the fleet you actually run, not one vendor's tool.

Over 100 commands: git operations, test runners, build tools (npm, cargo, and friends), docker, kubernetes, terraform, the AWS CLI, and more. Commands without a filter pass through untouched — output is never mangled. The dashboard's Opportunities screen shows you which high-volume commands are still passing through, so you know where the next savings are.

Nothing, from the developer's perspective. The agent is offline-first: filtering keeps working, metrics buffer locally (up to 30 days, then oldest-first drop), and reporting resumes with exponential backoff when the server returns. The developer's workflow is never blocked.

The open-source engine solves the problem for one developer. LazyToken adds what an organization needs: visibility (savings by team, in dollars), control (enforced install, signed central policy, managed updates), compliance (self-hosted, audit log, zero external telemetry), the Context Firewall DLP layer, AI FinOps, and accountability (support, SLA, one owner). The engine being open is a feature — your security team can audit every line that runs on developer machines.

No. The server is self-hosted in your environment — the vendor has no tenant, no copy, and no access. In air-gapped mode the server makes zero outbound calls, and that posture is technically enforced with a Kubernetes NetworkPolicy you can verify yourself.

Never. Redaction happens locally, in memory, before the output enters the model's context. What gets recorded is {type, tool, timestamp, device} — e.g. "an aws-key pattern was blocked in cat output" — never the matched value, the file path, or the command arguments.

No. Policies are signed with the org's Ed25519 key, and the agent verifies the signature against a key pinned at enrollment before parsing or applying anything. An invalid signature means the agent keeps the last good policy. Agent updates get the same treatment: sha256 + Ed25519 verification before an atomic replace.

It doesn't have to be. LazyToken ships a built-in anonymized mode: the dashboard shows teams and "Dev #N" instead of names. The choice is yours, set at policy level and recorded in the audit log — managerial visibility without personal surveillance.

Only from your internal server — never the open internet. ltk update downloads from your server's release channel and verifies both the published sha256 and an Ed25519 signature before an atomic self-replace that leaves the old binary intact on any failure. You can also roll out through your MDM.

The product targets under 60 minutes from server install to a dashboard with real data. Docker Compose for pilots, Helm for production, air-gapped runbook for the strictest environments — plus MDM scripts (Intune, Jamf, GPO) for the fleet. See the docs.