Argus — Testing for Coding Agents (Codex plugin)
Argus helps coding agents test the apps they build: exercise real user journeys, find bugs, and verify them with evidence. This plugin connects Codex to the local Argus MCP server with its 29-tool web profile, which covers exploration, interaction, verification, findings, reports, regression checks, and Playwright test export.
This is a project-maintained plugin, not an OpenAI product, and it is not listed in OpenAI's plugin directory.
Requirements and setup
Install uv and make
uv available to Codex. The launcher selects Python 3.12 and runs the exact
published argus-testing==0.12.0 release from the bundled uv.lock, which pins
dependency versions, sources, and hashes, in a virtual environment under the
installed plugin directory.
Before the first session, warm uv's cache and install the matching Chromium browser. This works from any directory:
uvx --python 3.12 --from argus-testing==0.12.0 --with playwright==1.64.1 python -m playwright install chromium
The warm-up pins Argus and Playwright; uv resolves the other dependencies afresh
rather than from the plugin's uv.lock, so they usually, but not always, match
it. A warm cache therefore usually removes most of the first-start download.
Linux may also need Playwright's system dependencies; see the
installation guide.
Add this repository's marketplace and install the plugin:
codex plugin marketplace add chriswu727/argus
codex plugin add argus-testing@argus
Plugins work in the Codex CLI and the ChatGPT desktop app; the Codex IDE extension
does not load plugins. Do not also configure Argus with codex mcp add: use either
the plugin or a manual MCP entry. uvx --from argus-testing argus-mcp --doctor
checks the browser and report directory if the first session fails.
Test an application
The bundled argus-test skill guides a session: it agrees the scope and asks
before anything with real-world effects, checks the environment, tests, rechecks
symptoms on fresh loads, and reports findings grouped as verified, observed, not
reproduced, or inconclusive. Ask Codex to test an application you are authorized
to exercise, for example:
"Test http://localhost:3000. Check saving, state after refresh, invalid input, and
cross-page consistency. Report reproducible issues with evidence."
The server supplies its testing protocol through start_session. The agent must
respect the user's scope and obtain authorization before consequential actions.
A completed tool call is not proof that an application passed. Independent
verification, agent observations, non-reproduction, and inconclusive checks remain
distinct. The plugin does not enable native macOS screen control, arbitrary page
JavaScript, or the optional monitoring port.
Privacy Policy
The Argus privacy policy covers data collection, storage, sharing, and retention. Argus has no operated cloud service or telemetry collector and does not call a model itself. Codex and its model provider receive the tool inputs and evidence included in a conversation, subject to their own policies. Reports, screenshots, journals, and state capsules stay on the user's computer until removed and may contain private application data.
Licensed under MIT.