Agent Sessions
A local-first macOS app that unifies browsing, search, and resume across 15 AI coding-agent session histories, with live per-session quota burn tracking for Codex and Claude.
Evidence shows read-only local session directory access, no telemetry, and network limited to optional Sparkle checks and a read-only price fetch — a reasonable scope; least_privilege misses full marks because docs/security.md is referenced but not in the sample to corroborate; user_confirmation gets 2 since explicit actions (terminal resume, cleanup workflows) are described but confirmation flows are not detailed; dependency_security gets only 1: Sparkle signing/notarization is mentioned but no dependency inventory or vulnerability management evidence; rollback gets 1: saved-session restore is mentioned but app-update rollback or data-change recovery is undocumented; source_attribution gets 2: a named-steward mechanism exists in STEWARDS.md, but the roster itself is outside the sample.
The README explicitly lists unverified or partially verified capabilities (Qwen resume untested end-to-end, fx resume untested, OpenClaw resume unsupported), showing good self-consistency; failure_messages gets 2: honest 'can't verify'/'no limit' states are described, but the three silent failures fixed in 5.1.1 are direct evidence that failures historically went unannounced; dependency_availability gets 1: CI runs build and test steps, but there is no evidence of actual pass status or a test-count baseline.
audience_and_scenarios gets 3: multi-agent power users are clearly targeted with concrete scenarios (quota monitoring, history recovery, resume); capability_boundaries gets 3: per-agent support level, image extraction gaps, and archive limits are enumerated — rare candor; trigger_precision gets 2: resume commands are generated per agent but some are unverified end-to-end; environment_fit gets 2: macOS 14+ stated, but assumptions around Sparkle updates and Homebrew distribution are not fully expanded.
information_architecture gets 3: documentation structure (guides, PRIVACY, security, CHANGELOG, summaries, CONTRIBUTING, STEWARDS) is complete and coherent; install_notes gets 3: DMG, Homebrew, and Sparkle paths are all documented; known_limitations gets 3: unsupported image extraction, untested resumes, and missing archives are proactively listed; versioning_changelog gets 3: the 5.1.1 fix list is detailed and links to the changelog; license gets 3: full MIT text present; maintenance_responsibility gets 3: the Maintained/Steward-verified/Steward-wanted model with a roster is an explicit accountability design; naming_stability gets 2: agent names and bundle id are consistent but no deprecation policy; examples_and_faq gets 2: screenshots and quota setup guidance exist but no FAQ and examples are illustrative images.
output_usability gets 2: search, browse, and resume outputs are clearly described but usability is unverified without execution; marginal_value gets 3: per-session quota attribution and multi-agent history aggregation are capabilities CLIs do not provide; cost_benefit gets 3: a free local MIT app that works on install.
claim_traceability gets 2: most claims carry links (verification report, changelog, STEWARDS.md), but link targets are outside this sample and unverifiable here; cross_source_corroboration gets 2: CI config matches README test commands and Python probe tests match Cursor metadata claims, but Swift-side code is absent from the sample; fact_inference_separation gets 3: the README cleanly separates 'verified', 'end-to-end untested', and 'steward-verified' states.
- Publisher is unverified by the registry; identity is treated as unknown and only repository content was assessed.
- Dependency-security evidence is thin: no dependency inventory or vulnerability-management notes, only mention of signed/notarized Sparkle updates; audit bundled dependencies before use.
- Several agents' resume is unverified end-to-end (Qwen, fx); Qwen archived history is browse-only and OpenClaw resume is unsupported — do not assume every resume button works.
- The three silent failures fixed in 5.1.1 are evidence that failures previously went unannounced; after upgrading, actively verify that search and cost figures behave correctly.
- The model-price list comes from an optional remote read-only fetch; cost estimates are approximations, not billing data.
- This is a static source review with low confidence; no execution or independent testing was performed.
What does this agent do, and when should you use it?
Agent Sessions is an open-source (MIT), local-first macOS app that consolidates AI coding-agent session histories scattered across your disk into one searchable interface. It supports fifteen session sources, including Codex CLI/Desktop/VS Code, Claude Code, OpenCode, Cursor Agent, GitHub Copilot CLI, Hermes, OpenClaw, Antigravity, Pi, Kimi Code, Grok CLI, Qwen Code, Devin CLI, and fx. Core features include unified search, transcript inspection, image browsing, saved-session recovery, and one-click resume commands for supported CLIs (e.g., devin --resume <id>). Its Quota Meter attributes Codex and Claude 5-hour and weekly limit burn to the individual session in real time, with percent, tokens/hour, and dollar lenses — priced per model rather than blended. All data is indexed locally with no telemetry; the only network activity is an optional Sparkle update check and an optional read-only fetch of a public model-price list. The app requires macOS 14+ and installs via DMG or Homebrew Cask.
The app scans local agent session directories in read-only mode (e.g., ~/.codex/sessions, ~/.claude/projects, ~/.gemini/antigravity/brain, ~/.copilot/session-state, ~/.cursor/projects, ~/.hermes/state.db, ~/.local/share/opencode/opencode.db, ~/.factory/sessions, ~/.openclaw/agents, ~/.pi/agent/sessions, ~/.fx/sessions, and Devin's shared SQLite sessions.db) and builds local indexes for search and navigation. You can browse, filter, and full-text search transcripts across all sources, inspect tool calls/outputs, and browse images from sessions; for supported CLIs, right-click copies the exact resume command or opens it in Terminal.app, iTerm2, or Warp. The Quota Meter shows live per-session burn against Codex and Claude 5-hour and weekly windows, switchable between 5-hour, weekly, tokens/hour, and dollar lenses, with the dollar lens pricing each model in a session at its own rate. Since 5.0, agent support is implemented as plug-in adapters, with a documented recipe for contributing new sources.
- A developer running several coding agents at once who needs to know which individual session is eating the 5-hour or weekly quota, not just an account-wide percentage
- Recovering a solution or snippet from a Codex or Claude Code session days ago, when local history is too large to search manually
- A team using multiple agents (Cursor, Copilot CLI, OpenCode, etc.) that wants unified search and filtering across all session records
- Resuming an interrupted agent session by copying the correct CLI resume command (e.g., devin --resume <id>) with one click
- Cost-conscious users mixing Opus orchestrators with Sonnet subagents who need per-model pricing instead of a blended number
- Privacy-sensitive users who want fully local indexing with no telemetry and no session data leaving the machine
What are this agent's strengths and limitations?
- Attributes quota burn to individual sessions live, and the dollar lens prices each model in a session at its own rate (an Opus orchestrator driving Sonnet subagents is costed per model) — something an ordinary quota meter cannot do
- Covers fifteen session sources with a plug-in adapter architecture since 5.0; the contribution recipe has been proven by an outside contributor who added Devin CLI and fx
- Strict privacy design: local-only indexing, no telemetry, read-only access to session directories, documented in docs/PRIVACY.md and docs/security.md
- Resume commands are end-to-end verified for supported CLIs (e.g., Devin verified on 3000.5.20), with three terminal launch targets
- macOS 14+ only — Windows and Linux users cannot use it
- Per-agent gaps remain: Qwen and fx image extraction is unsupported, Qwen resume is end-to-end untested due to authentication, OpenClaw resume is not supported, and Devin keeps only active sessions with no archived history
- Eleven of fourteen agent sources are marked 'Steward wanted', meaning nobody tracks format changes — upstream CLI updates can break parsing
- Depends on each agent's local storage format; changes can fail silently (e.g., 5.1.1 fixed Claude sessions being unsearchable since a date with no warning), so staying updated matters
How do you install or deploy this agent?
Option A (DMG): download AgentSessions-5.1.1.dmg from GitHub Releases and drag Agent Sessions.app into Applications. Option B (Homebrew): brew tap jazzyalex/agent-sessions && brew install --cask agent-sessions. Requires macOS 14+. Automatic updates come via Sparkle (signed and notarized).
How do you use this agent?
After installing, launch the app; it reads the supported default CLI session directories and builds local indexes (re-indexing after upgrades is automatic — nothing to do by hand). Browse, filter, or full-text search transcripts in the unified session list; use in-session Find to jump to relevant tool calls and outputs; right-click a resume-capable session and choose Copy Resume Command or Resume in Terminal.app, iTerm2, or Warp. For the Quota Meter, enable live session detection in settings and run your agents in a terminal or Codex/Claude Desktop; for best session rows, set the terminal window title to the repo name and name each tab/session clearly. Show or hide the Quota Meter from the View menu (⌘⌥⇧C) and drag it anywhere on screen.
How does this agent compare with similar options?
The source material names no direct competitors; the README contrasts its Quota Meter with an 'ordinary quota meter' (account-wide percentage only) without naming a product.