Cut your coding agent's token costs by up to 90%.
AgentLogs analyzes your agent session logs and turns them into artifacts and context that make your coding agents run leaner — fewer loops, faster error fixes, and far lower token bills.
curl -fsSL https://agentlogs.promptingco.com/install.sh | shEvery signal stays grounded in evidence
AgentLogs runs a small, auditable kernel. Raw events flow into scoped views, get reduced into typed signals, and surface as findings — and every step points back to the events that produced it.
- 01Source~/.codex/sessions
- 02Artifactone session
- 03Eventmessages, commands, edits
- 04Segmentscoped windows
- 05ProjectioncompactTimeline, failedCommands
- 06Extractorlocal rules or LLM
- 07Signaltyped + evidence
- 08Findinghuman insight
- 09Reporttext, JSON, Markdown
what signal logic runs
where extraction executes
how it talks to a model
Extractor identity is separate from backend identity. The same verification-status signal runs through local rules, a local Ollama model, or a remote backend — without changing the raw evidence model.
Audit agent runs, don't re-read them
Built around a signal-analysis kernel that keeps structured observations grounded in raw, auditable evidence.
Local-first by default
01Discovers and analyzes sessions on your machine. There's no hosted service to sign up for and no requirement to send raw logs anywhere.
Evidence on every signal
02Each signal carries pointers back to the exact events — session id, event index, command output — that justify it. No claim without proof.
Map / reduce on long sessions
03Long transcripts are split into deterministic windows, projected into focused views, then folded into session-level signals — not stuffed into one prompt.
Backend-neutral extraction
04Run the same signal definition through local rules, a local model, or a remote backend. Extractor identity is decoupled from where it executes.
Reproducible reports
05Provenance records the extractor version, prompt version, model, and config hash, so a report can be regenerated from a frozen analysis run.
Codex today, more to come
06The first source adapter reads Codex session history. The shared kernel is domain-neutral, ready for other coding agents without reinventing the vocabulary.
Evidence answers “what supports this?” — provenance answers “how was it produced?”
Every signal is structured data with both — so results are auditable and reproducible, never a black box.
- --since
- 7d, 24h — time window to scan
- --session
- latest or a session id
- --limit
- cap the number of sessions
- --format
- text or json output
- --backend
- where extraction runs
{ "signalId": "verification-status", "scope": "artifact", "value": "verified_after_final_edit", "confidence": 0.91, "evidence": [ { "eventIndex": 55, "snippet": "apply_patch bar.ts" }, { "eventIndex": 72, "snippet": "bun test exited 0" } ], "provenance": { "extractor": "verification-status", "extractorVersion": 1, "model": "qwen2.5-coder:14b", "configHash": "sha256:…" } }
Install AgentLogs
Free and local. Point it at your Codex sessions and read your first report in seconds.
$ curl -fsSL https://agentlogs.promptingco.com/install.sh | sh$ irm https://agentlogs.promptingco.com/install.ps1 | iex$ go install github.com/promptingcompany/agentlogs@latest