Show HN: TMA1 — local-first observability for LLM agents. Tokens, cost, tool calls, latency, full session replays; stored locally and queryable with SQL. Open-source (JS, 45★). Read: https://tma1.ai/ Take: if an agent acts for you, you must be able to prove what it did.
Why local-first: privacy + auditability. Vendor dashboards often mask telemetry and ship data offsite. TMA1 keeps traces on your host so compliance teams can run SQL audits and replay sessions without sending client data to a third party.
How I'd use it: run TMA1 as a telemetry sidecar. Emit events with session_id, tool_name, token_count, latency, and result. Staging = full payloads, Prod = metadata + sampled replays. Query for avg tokens/session, tool-failure-rate, and 95th‑percentile latency.
Tradeoffs and next step: you get control but inherit ops — retention, backups, indexing, access controls. Start small: wire one agent flow, add volume-based sampling and a 30/90d retention policy, then ask compliance: “What exact proof do you need?”
Top comments (0)