Agent Sessions (macOS)

Live per-session quota burn for Codex and Claude — see which session is eating your 5-hour and weekly limits, priced per model. Plus searchable transcripts across Codex, Claude, OpenCode, Cursor, GitHub Copilot CLI, Pi, Kimi Code, Grok CLI, Qwen Code, Devin CLI, fx, Antigravity CLI, Hermes, and OpenClaw, with image browsing and one-click resume where the source supports them. macOS, local-only.

Session-Bench: how ten agents' session formats score — 20 pass/fail gates, evidence behind every cell.

If Agent Sessions hands you back something you thought you'd lost, star the repo. It is the only thing this project asks for.

• Requires: macOS 14+
• License: MIT
• Security & Privacy: Local-only. No telemetry. Details: docs/PRIVACY.md and docs/security.md

New in 5.0 — agents are now plug-in adapters, Qwen Code joins as the thirteenth source, and adding the agent you use is a documented recipe (26 shared files before, 12 now). See what's new ↓

Overview

Run three agents at once and a normal quota meter tells you "60% used" — not which one spent it. Agent Sessions attributes burn to the individual session, live, against your Codex and Claude 5-hour and weekly windows. Pick the lens you want (5-hour, weekly, tokens/hour, or dollars); the $ lens prices each model in a session at its own rate, so an Opus orchestrator driving Sonnet subagents is costed per model instead of blended into one number.

It's also a local-first Mac app for finding useful work coding agents already wrote to disk — Codex, Claude, OpenCode, Cursor Agent, Hermes, OpenClaw, Antigravity, GitHub Copilot CLI, Pi, Kimi Code, Grok CLI, Qwen Code, Devin CLI, and fx histories in one searchable view, with transcript inspection, image browsing, saved-session recovery, and resume commands for supported CLIs.

Security & Privacy

• Local-first: session data stays on your Mac.
• No telemetry, analytics, remote logging, advertising identifiers, or session-history uploads.
• Reads local session folders you choose, plus supported default CLI locations.
• Builds local indexes/databases for search and navigation.
• Explicit actions may open Terminal/iTerm2 resume commands or run status/probe cleanup workflows.
• The only network activity is optional Sparkle update checks and an optional read-only fetch of a public model-price list (for the runway's $ estimate) — neither sends any personal or session data.

Details: docs/PRIVACY.md and docs/security.md.

What's New in 5.1

TL;DR - Two more coding agents: Devin CLI and fx (vercel-labs). Both land in the same searchable list as the rest, each with a Settings pane and a resume command.

New in 5.1:

• Devin CLI, the fourteenth agent source — browsing, search, filtering, Analytics and the transcript view over Devin's shared SQLite store. Devin keeps every retry and edit as a branch, so a session shows the live conversation rather than every path it took to get there. Settings adds binary and storage-root controls, and Copy Resume Command produces devin --resume <id>. Devin keeps only active sessions, so there is no archived history to browse, and image extraction is not yet supported.
• fx (vercel-labs), the fifteenth — one directory per session under ~/.fx/sessions, with every kind of turn rendered: ordinary replies and their narrated tool steps, background commands and where their output went, the summary that replaces auto-compacted history, and interrupted turns with whatever finished before the cut. Text fx stored as raw bytes rather than UTF-8 decodes back to readable output instead of disappearing.
• Both agents are contributed, not ported — @thedavidweng added them against the documented recipe from 5.0, which is the first real test of whether that recipe works for someone outside the project. Both stay steward wanted until someone running the CLI confirms resume end to end; the support matrix says so on every surface.

Recent releases — 5.0: agents became plug-in adapters and Qwen Code joined as the thirteenth source. 4.8: Grok CLI joins as the eleventh agent source, and Analytics counts every agent you have enabled. 4.7: Kimi Code joins as the tenth source, and active Claude cloud sessions appear in the Quota Meter. 4.6.4: Compact and Full Agent Cockpit retired, leaving the Quota Meter as the only mode. 4.6: paste-a-cookie Claude web usage, no CLI or Full Disk Access needed. Full history in the changelog.

Core Features

• Browse and search Codex CLI, Codex Desktop, and Codex VS Code sessions in one place.
• Browse Claude CLI and Claude Desktop sessions with consistent labels and project context.
• Browse Cursor Agent transcripts from Cursor's local storage, enriched with Cursor chat metadata when available.
• Hermes Agent sessions participate in browsing, search, filtering, analytics, and resume workflows, including current ~/.hermes/state.db storage.
• OpenClaw sessions participate in browsing, search, filtering, image browsing, and deleted-session visibility while ignoring trajectory traces. OpenClaw resume is not supported.
• Pi CLI sessions now participate in browsing, search, filtering, and resume workflows.
• Qwen Code sessions (new in 5.0) participate in browsing, search, filtering, and Analytics, including active-chain tool calls and reasoning recorded in local JSONL history. Active-chat resume actions are implemented from the installed CLI's help and reader behavior, but remain end-to-end untested because authentication blocked a disposable 0.21.13 run; archived Qwen histories remain browse-only. Qwen image extraction is not yet supported.
• Devin CLI sessions (new in 5.1) participate in browsing, search, filtering, Analytics, and resume workflows from the shared SQLite sessions.db under the CLI data directory, with working directories read from the sessions table's working_directory column. Devin keeps only active sessions in that database, so there is no archived-history surface; the records carry inline base64 images, but image extraction is not yet supported. Resume command plans are tested against probe-advertised cap