All your agents.Every trace, every eval.On your machine.

Local-first observability for AI coding agents. No cloud. No SaaS. Just your filesystem.

$brew install duncankmckinnon/tap/thirdeye

terminal

$ thirdeye list --since 2026-05-01

Platform	Started	Events	Tags	Session ID
claude	2026-05-22 14:08	47	review,bug	sess_3f9e2a
cursor	2026-05-22 09:31	23	-	sess_b8c402
codex	2026-05-21 18:55	91	refactor	sess_71d9aa
gemini	2026-05-20 11:14	12	-	sess_4e0f6b
claude	2026-05-19 22:02	56	migration	sess_a2c1d3

$ thirdeye search "migration" --tag review --platform claude

2 result(s):

sess_3f9e2a · 2026-05-22 · "fix migration drift in user table" → seq 12, 31

sess_a2c1d3 · 2026-05-19 · "plan migration rollout for staging" → seq 4

━━━ Summary ━━━

5 sessions across 4 platforms · 229 events captured

How it works

Install

One command. Mac and Linux. No background daemon.

$ brew install duncankmckinnon/tap/thirdeye

Wire it up

Attach to every agent you use. Tracing starts on the next session — no code changes.

$ thirdeye add --claude --cursor --codex --gemini --copilot

Read your history

List, tail, search, tag, and evaluate every session that ever ran on this machine.

$ thirdeye list && thirdeye search "auth refactor" --tag review

From session to insight in two commands

Search and inspect

$ thirdeye search "OAuth bug" --tag review

3 result(s) — claude (2), codex (1)

$ thirdeye events sess_3f9e2a

[seq 12] tool_use Edit auth/oauth.py

[seq 13] tool_result ok

[seq 14] message user "still failing on refresh"

[seq 31] tool_use Bash pytest tests/test_oauth.py

Evaluate a session

$ thirdeye eval run sess_3f9e2a \

--agent claude --using token-efficiency

→ dispatching claude in read-only mode...

✓ completed in 41s

VERDICT: warn

findings:

seq 12 redundant Edit before reading file

seq 31 test rerun could have been targeted

Everything you need to see what your agents did

Multi-agent capture

Claude Code, Cursor, Codex, Gemini, Copilot — one unified history.

Local-first storage

Everything stays in ~/.thirdeye/. No cloud, no telemetry, no account.

Per-turn token usage

JSONL sidecars + a SQLite index. Roll up by model, time, or platform.

Tag & search

Annotate any event with tags. Search substring or filter by platform, tag, or date.

Evals as rubrics

Named directive rubrics graded by claude/codex/gemini as LLM-as-judge.

Agent-friendly

Skills ship with thirdeye so your agents can read their own history.

Try thirdeye

Install in 60 seconds. Start tracing your next session.

$brew install duncankmckinnon/tap/thirdeye

Get Started

Pair with workbench

Dispatch parallel agents and watch the trail unfold.

Visit wbcli.com