Persona installation is the useful distinction
Agency Agents App puts a desktop interface around installing persona files into coding tools. The separate agency-agents repository supplies the catalog; Claude Code, Codex and the other destination tools execute the resulting instructions. Choose this app for file management rather than agent orchestration.
The supplied README lists eight install targets: Claude Code, Codex, Gemini CLI, GitHub Copilot, Qwen Code, Cursor, opencode and Osaurus. Its table gives Cursor and opencode project scope and the other listed targets user scope. That table should not be read as an exhaustive release manifest: the v0.3.2 release notes also announce DeepSeek Harness support. Check the exact app version and destination before deploying. Aider, Windsurf, OpenClaw, Antigravity and Kimi remain recognized-only in the supplied README.
According to the README, the local ledger records the source hash, rendered hash, tool, destination, scope and relevant project path. Reconciliation compares rendered source with installed bytes and classifies files as current, outdated, modified, removed or foreign. The documented backup behavior protects modified files before destructive operations. These are documented capabilities, not desktop workflows we verified.
What our recorded run establishes
Our run was recorded on October 6, 2026 at commit 2d6bbfe8e15e65a4c18bd3228652ccd8ceeae0c0. Dependency installation exited successfully after 12.613 seconds, reported as 13 seconds in the stored summary. The raw output says npm added 62 packages. The separate dependency metric is 69 packages and 108 MB; it does not mean npm added 69 packages.
The build step exited successfully after 22.545 seconds, reported as 23 seconds in the summary. Its output shows SvelteKit generating a static site into the build directory. It also reports chunks larger than 500 kB after minification. This is evidence of a successful frontend build, not a native desktop package build or an installed-app check. We have no recorded launch, persona installation, reconciliation or coding-tool workload from this run.
The recorded sandbox configuration was 3 CPUs, 8 GB of memory and image lab-node:22. Detection recorded Node major 22, but no separate runtime-version output was supplied. Repository metrics recorded 598 files, approximately 39,432 source lines and 15.9 MB. The detector found two CI workflows, no tests directory and no Dockerfile. These detector results do not establish the absence of Rust unit tests or other checks.
The test step was skipped with reason "no script". The README documents npm run check, cargo test --manifest-path src-tauri/Cargo.toml --lib, npm run build:phase-c and npm run build:phase-c:full. The harness did not run those checks. A missing npm test script is therefore a harness-discovery limit, not evidence that the project has no tests or that tests failed.
npm audit recorded 7 advisories: 4 high and 3 low, with zero critical or moderate findings. The number of scanned packages and affected runtime paths are unknown in the supplied record. The totals warrant dependency review but do not prove that the desktop application is exploitable.
What v0.3.2 fixed, and who verified it
The v0.3.2 release page displays an October 3 publication timestamp; its release heading is dated October 2, 2026. The maintainer describes renderer drift since mid-August and fixes for tool names, YAML descriptions, horizontal rules and install paths.
For Qwen, the notes say 17 agents could edit existing files but could not read or create files and had no shell because their tool grants used Claude Code names. Three agents had invalid YAML descriptions, and 137 agent bodies lost horizontal rules. These are historical defects described as corrected in v0.3.2, not failures reproduced in our run.
The release reports 2,538 identical byte comparisons across 282 agents and nine tools against catalog commit d3f71c4, passing cargo tests, zero npm run check errors, and successful Linux and Windows release builds. Those are maintainer-reported release checks. They are separate from our frontend build and are not our passing-test count or independent proof of exhaustive compatibility.
For a rollout, install a small selection into the exact destination tool, inspect the generated files and exercise their intended behavior. Record both the app version and catalog revision so later changes can be traced.
Packaging and local operation
The v0.3.2 downloads list signed and notarized macOS packages for Apple Silicon and Intel, requiring macOS 13+. Linux packages target x86_64. Windows installers target x64 and ARM64; the release says they are unsigned and trigger a first-launch SmartScreen warning. An environment that requires signed Windows installers cannot treat that package as meeting its policy.
Issue 89 reports a blank, unstyled v0.3.0 window for the deb package under WSL2/WSLg. The reporter's follow-up says the native Windows x64 build worked on the same machine. This is a specific reported rendering problem, not evidence that the app fails generally on Windows or ordinary Linux desktops. Our run did not evaluate either rendering path.
The README says core browsing and install tracking are local and require no account. Catalog refresh, GitHub sign-in, optional GitHub features and update checks use explicit network paths controlled through settings. It also says GitHub tokens stay in the platform keychain and that no telemetry or product analytics are included. We did not independently inspect network traffic or token storage.
Choosing a route
The upstream Agency Agents catalog documents script-based installation of the same personas and manual copying or adaptation. That is the relevant alternative if you prefer terminal commands. Skill Mix addresses a related task: its README describes skill discovery, installation, scope management and updates across Cursor, Codex and Claude Code. Those descriptions do not establish equivalent persona rendering or tested compatibility.
Start with the desktop app if browsing and tracking persona files is your immediate need. Before expanding its use, check the generated output, your platform's packaging requirements and the dependency advisories recorded for the reviewed commit.
How to read the scores
These are editorial evidence ratings, not benchmark results. On the five-point scale, 1 means substantial documented obstacles, 2 means limited supporting evidence, 3 means a documented path with material validation gaps, 4 means clear documentation or directly observed success within the stated scope, and 5 requires broader corroborated validation than this record supplies. Community is judged from visible participation evidence, not assumed adoption. No overall score is calculated.

