The black box for your coding agent.
When Claude Code or Codex CLI loops, drifts off-goal, or quietly burns two million tokens, agentfdr shows you why — turn by turn, after the fact. And while twenty sessions are running, one kanban shows which of them is waiting on you.
Zero instrumentation · zero cloud · zero config — it reads the transcripts Claude Code and Codex already write.
The evidence is gone by the time you notice.
Autonomous agent failures are hard to debug — unless someone kept the flight recording.
Edit → test fails → same edit, for 40 minutes.
You asked for a bugfix; it refactored the router.
A huge tool result crowds the context, cache stops hitting, every turn re-reads 200k tokens.
“Done!” with failing tests — or no stop at all.
Working with agents is becoming loop engineering — designing and operating the loop your harness runs. You can't engineer a loop you can't see. agentfdr is the instrument panel.
The crash investigator's toolkit.
The whole session on one screen: every turn's tool calls, context-window composition, and output tokens, with prompt and compaction markers.
Tool loops, error streaks, context bloat, token spikes, cache thrash, file churn, and refusals — flagged automatically, one click from the evidence.
A resizable side panel with the usage breakdown and every tool call's duration, result size, and snippet. Step through turns with ←/→.
agentfdr watch follows a session that's still running — watch the context climb in real time.
agentfdr assert --no-loops --max-tokens 2M exits 1 on violation. Put a tripwire on your autonomous pipelines.
agentfdr blame renders the whole analysis as markdown, ready to paste into an issue or a Slack thread.
agentfdr search scans every prompt, reply, tool call and result across all your sessions — and the Search tab jumps straight to the matching turn.
agentfdr diff puts the failed attempt next to the successful retry: stats, anomalies, tool mix, and which files each one touched.
agentfdr board puts every running Claude Code session on one kanban — needs you / working / idle — with the current task, cost so far, and a warning when two sessions touch the same file. Send instructions, approve or stop tmux sessions from the page.
The control tower for parallel sessions.
One session needs a flight recorder. Twenty need a control tower: which one is waiting on you, which two are editing the same file, how much of each account's plan is left. agentfdr board answers that on one screen — and lets you answer back.
Needs you (permission prompt, question, plan approval) · working · idle · recently ended — CLI and the Claude desktop app side by side, each card with the task you typed, what the agent is doing right now, cost since that prompt, and the files it touched. Group by state, project, account, or your own drag-and-drop lanes.
tmux-hosted sessions take instructions, approve / deny a permission prompt, or stop a turn — two-click, never modal, refused if the session's state changed since the card was drawn. Start a new session in any repository from the same page.
Sessions on the same repository are grouped (worktrees resolve to their parent) and a banner names the files more than one live session is editing. Per-account 5-hour / 7-day plan limits sit at the top, straight from Claude Code's own status line.
Reads the live-process registry and the tail of each transcript that Claude Code already keeps — no hooks, no daemon, milliseconds per refresh, loopback only.
Know what your plan is really burning.
agentfdr usage aggregates every project's transcripts into the same shape your subscription is metered in: the current 5-hour window, daily history, and the rolling week — with a per-model breakdown and estimated cost.
Your plan tier is auto-detected. Set budgets, calibrate them once against Claude Code's /usage screen, and get warning bars before you hit the wall.
One command away.
agentfdr # open the newest session's timeline agentfdr list # all recorded sessions across projects agentfdr watch # live: follow the running session agentfdr blame 35cb18 # markdown autopsy for an issue agentfdr board # every running session on one kanban; steer, launch agentfdr usage # plan burn: 5h window / daily / weekly agentfdr assert --no-loops --max-tokens 2M # CI gate