Persistent agents can turn computational spending into background noise. A long-running system may be productive, wasteful, or both, but without telemetry the difference is difficult to see until after the bill arrives.
A TypeScript, SQLite, and dashboard tool that collects and summarizes cost telemetry from an OpenClaw-style agent environment.
The aim. Make agent operating costs visible enough to budget, compare, and debug.
01The problem behind the project
Persistent agents can consume resources invisibly. Cost telemetry makes experiments easier to evaluate and reduces the chance of discovering runaway usage after the fact.
CostClaw was designed as an observability layer for that problem. Instead of treating token usage as an incidental provider statistic, it organizes cost around agent work so an operator can compare runs, notice unusual consumption, and ask whether autonomy is economically justified.
People operating autonomous or long-running agents may benefit. Logs can contain sensitive prompts or identifiers, so public analytics must be scrubbed or synthetic.
02How it took shape
A local plugin architecture, SQLite storage, collection hooks, summaries, and a web dashboard.
The implementation used TypeScript, SQLite, collection hooks, summaries, and a local dashboard inside an OpenClaw-style environment. A v0.1 release and public discussion demonstrated interest in the problem, though not durable compatibility or verified savings.
Josiah identified the cost-visibility problem, set the product direction, evaluated the tool in practice, and used coding agents for implementation.
A v0.1 implementation shipped in 2026 and its launch discussion received substantial attention, but the current repository and integrations have not yet been revalidated.
03What the project means now
Its present value depends on an audit. Agent platforms and pricing models change quickly, so a historical telemetry plugin can look functional while reporting the wrong thing. If CostClaw is now obsolete, the honest article still records an early attempt to make agent economics observable.
The tool may be obsolete as agent platforms change, and known repository issues have not been audited. No private logs or credentials belong in the public case study.
Agent infrastructure ages quickly; an honest portfolio record has to distinguish historical usefulness from current compatibility.
Audit the repository with a modern coding agent, fix viable paths, and archive it clearly if current platforms have made it redundant.