AgentOps is free today. Everything documented here — the CLI, hooks, scoring, the leaderboard, built-in and community plugins — costs nothing. A pro tier is coming, which will include premium plugins and team features. The core local-first tool stays free.
By default: nothing. AgentOps is local-first — telemetry lives in a SQLite database inside .agentops/ in your project, and raw command output stays on your disk.
When you run agentops submit, numbers are sent: your score, grade, the five dimension scores, token and command counts, cache stats, and the six Communication Efficiency Score integers. The server rejects any submission where those fields are not numbers in the expected ranges. The full field list, the validation code, and the automated leak test are documented in Model Communication and Privacy.
No. Agent transcripts are read locally to compute CES (six integers, 0–100) and never transmitted. This is enforced by a canary-based leak test in CI, not just by policy: tests/ces-submit.test.ts plants marker strings in a synthetic transcript and asserts none survive into the submit payload.
The one exception is the opt-out data pipeline, which can include outputs of commands your agent ran (never transcripts, and with secrets automatically stripped) to improve plugin quality. The CLI announces this explicitly before it ever happens, and you can turn it off by setting telemetry.enabled: false in .agentops/config.yaml.
Today:
agentops hooks install claudeagentops hooks install codexagentops hooks install gemini (requires Gemini CLI v0.26.0+)agentops opencode installCursor, Windsurf, and Aider support is planned. Any tool that can call a shell hook or run commands through agentops run <command...> works in a degraded manual mode already.
A weighted sum of five dimensions: savings rate (0.35), cache efficiency (0.25), waste control (0.20), summary coverage (0.15), and volume bonus (0.05). The server recalculates every submitted score and rejects mismatches. Full formulas and the grade table: How Scoring Works.
No. Once you are logged in (agentops login), AgentOps auto-syncs sessions in the background from the hook path. Sync is deliberately conservative: it only fires for sessions with at least 5 commands, at most once per 10 minutes, or earlier if 20 new commands have accumulated. It runs detached and silent — it can never slow down or break your agent session. If you are not logged in, nothing syncs. agentops submit remains available for explicit, confirmed submissions.
Nothing visible. Every integration point is fail-open: malformed hook payloads, a missing transcript, a broken plugin, or a local error all degrade silently and your agent session continues. Run agentops doctor and agentops hooks status to diagnose afterward.
Run:
agentops --version
agentops doctor
agentops hooks status
agentops statusExpected success: version prints, doctor reports a healthy local workspace, hooks show OK, and status shows your player dashboard after login. If status says you are not logged in, run agentops login. If hooks are missing, rerun the agent-specific command from Getting Started.
Use this order:
agentops doctor
agentops hooks status
agentops hooks install claude # or codex/gemini
agentops opencode install # only for OpenCode
agentops commands --last 5If there are no recent commands after a real agent session, the host integration is not invoking the wrapper. Reinstall the relevant hook and verify from the same project root.
Yes. Set the AGENTOPS_TOKEN environment variable and the background sync uses it instead of the on-disk auth token.
Near-term, in order of priority:
Follow development at github.com/ordemri/agentops.
Getting Started takes you from install to a scored session in about 5 minutes. Then check your dashboard.