# Compatibility: one package, different host capabilities

Official installation sources rechecked for 1.6.0-rc.3 on 2026-09-10. No live GUI import was performed here.

| Surface | Package path | What must be verified separately |
|---|---|---|
| Claude Desktop Cowork with custom plugins | Upload this plugin ZIP using the custom-plugin flow | Plugin enabled, skills load, authorized real host folder, Local Bridge, local scheduled runs |
| ChatGPT desktop Work/Codex with plugin support | Extracted plugin via documented local/repo marketplace | Plan/admin eligibility, local file tools, actual installed paths, host scheduler/folder behavior |
| Single-Skill upload interface | Use exported per-skill ZIPs, NOT this multi-skill plugin ZIP | All dependencies installed and reachable |
| Ordinary cloud chat / generic reminders | Reference/review only unless actual authorized host tools are supplied | Does not imply access to a Mac path, Keychain or Local Bridge |

Portable root plugin.json plus .claude-plugin/plugin.json and .codex-plugin/plugin.json describe
the same skills/version. There are no MCP server declarations, remote endpoints, automatic hooks,
background jobs or permission-changing code in this distribution.

## Documentation conflicts are not solved by a confident guess

The fetched Claude scheduling page says cloud tasks cannot attach computer folders, while its
manual setup section also lists an optional folder and says local-file/app tasks run locally.
Treat these as surface/version-dependent or unresolved documentation, and verify the real host
UI/schema plus actual scheduled run. Do not convert a docs contradiction into an invented API.

The fetched OpenAI task page says a task created in a project cannot access uploaded/project-stored
files. That is not a promise of access to a local host folder; Work/local task capabilities need
separate proof. A generic cloud task cannot execute a local bridge protocol by naming a Mac path.

Custom schedules: the current Claude manual setup list has no Custom option. That does not prove
all API/legacy cron support was removed. Inspect supported fields, read saved timing and test it.
The native /schedule command, when available, is not proof of programmatic invocation or scope.

## Readiness is scoped, not universal

Installing this plugin does not install Outloop or its host daemon, grant provider access, connect
Chrome/Playwright, resolve missing Context or create a scheduler. No app's sandbox or admin policy
is overridden. The same runtime may be usable on one workstation and pending on another.
Only one assistant/host may own a given intake key; do not activate both Claude and ChatGPT for the
same work simultaneously. Cross-platform portability of files is not cross-platform runtime proof.

## Historical rc.3 host evidence — not rc.4 installation proof

Read-only application metadata confirmed Claude Desktop 1.49585.0 and an OpenAI desktop app
26.901.51231 (bundle identifier com.openai.codex, build 8109) on the build Mac. These identify the
observed apps, not an import result or plan/admin eligibility. Claude's local plugin manifest
validator passes without errors/warnings; session-scoped CLI discovery lists rc.3 and all eight
skills. This does not prove Cowork GUI import. No global plugin installation, app restart or live schedule was changed.

An isolated authorized GUI import, installed-byte read-back and fresh session are still required
for each advertised host. Actual provider, bridge, connected-folder and autonomous schedule proofs
remain separate acceptance actions. The coding session's approved Context read is administrative
transport evidence only; it does not prove execution by the new generated runtime.

The isolated OpenAI CLI marketplace discovery attempt returned an empty inventory; GUI import
remains NOT_PROVEN because Computer Use refused access to that app. Claude Desktop customization
controls were observed, but no isolated GUI install was completed while the user was interacting
with the app. These limits do not change the public package format or authorize live client tests.

Current installation references: [OpenAI plugin packaging](https://developers.openai.com/plugins/build/plugins)
and [Claude plugin installation](https://support.claude.com/en/articles/13837440-use-plugins-in-claude).

For rc.4, the package format is retained. No historical GUI, CLI discovery or screenshot is proof
of rc.4 installed execution. Current local validation is recorded in PROOF.md and qa/.