Agent Verify is a native CAPTCHA and human-verification solver for ChatGPT and Codex browser automation. It keeps long-running agent tasks moving through reCAPTCHA, hCaptcha, Cloudflare Turnstile, image grids, text CAPTCHAs, sliders, rotation puzzles, press-and-hold controls, passive bot checks, and refreshed challenges.
It uses the model, vision, transcription, and browser controls already attached to the task. There is no paid CAPTCHA service, solver API key, separate browser, proxy network, or bundled model download.
Agent Verify treats human verification as an interruption, not the end of the task. It remembers the exact step that was blocked, clears every immediate challenge round in the active tab, checks that the page unlocked, and continues the original task.
The plugin follows the approval state and scope already enforced by the ChatGPT or Codex browser. It doesn't add its own per-tile, per-click, or per-round prompts.
For ordinary pages, Agent Verify remembers the last completed step and the next action. When a site exposes a WebMCP tool, it preserves the interrupted tool call and its arguments. It retries only when the site reports that the action didn't run or provides an idempotency key or resume handle. If the result is unknown, it preserves the pending action instead of risking a duplicate booking, order, or form submission.
Agent Verify selects a solving route from the challenge it can see instead of relying on a provider allowlist:
- reCAPTCHA, hCaptcha, and Cloudflare Turnstile interactions.
- Checkbox and multi-round image CAPTCHAs.
- Distorted text CAPTCHAs using host vision or OCR.
- Provider audio when the host exposes native transcription.
- Sliders, drag targets, rotation puzzles, and press-and-hold controls.
- Passive browser checks that need time rather than clicks.
- Expired, refreshed, and reissued challenges.
After every change, Agent Verify observes the page again. It never reuses an answer, transcript, or coordinate map after the challenge changes.
These results came from Agent Verify controlling the active browser with the host model's native capabilities:
| Verification test | Result |
|---|---|
| Google reCAPTCHA official demo | Passed in the ChatGPT/Codex in-app browser and connected Chrome. |
| hCaptcha public demo | Passed a live two-round image challenge in both browsers. |
| Cloudflare Turnstile forced-interaction test | Passed in connected Chrome. The in-app browser rendered the widget but couldn't activate its checkbox. |
| Seven-route browser lab | Passed checkbox plus refreshed grids, distorted text, slider, rotation, press-and-hold, passive wait, and expiry/reissue in both browsers. |
| Long-task resume | Resumed the pending appointment with one booking write. The ambiguous post-commit test preserved the task with one call, one write, and no duplicate. |
The full test report separates deterministic browser coverage from live provider results. The test protocol explains how to repeat the local suite. Provider audio is supported through host-native transcription but hasn't completed a live provider test yet.
Add the tagged GitHub marketplace and install the plugin:
codex plugin marketplace add anotb/agent-verify --ref v0.2.0
codex plugin add agent-verify@agent-verifyRestart the ChatGPT desktop app or Codex, then start a new task.
Give ChatGPT or Codex the full browser task. Mention Agent Verify at the start if you expect a CAPTCHA, or tell the agent to use it when verification appears.
Book the earliest appointment before 11:00 AM. If a CAPTCHA or human-verification check appears, use Agent Verify to clear it and continue the same task.
Complete this checkout. If verification interrupts it, clear every immediate round and continue without repeating a completed order.
Agent Verify doesn't operate a server or send challenge material to a separate solver. Challenge images, provider audio, visible page state, and task-resume state stay inside the active ChatGPT or Codex workflow and the site you are already using. The plugin doesn't export browser cookies, signed-in session data, or clearance tokens.
Read the privacy policy, terms of use, or open a GitHub issue.
Agent Verify needs a browser surface that can display and operate the challenge. An explicit access denial, rate limit, unsupported-browser message, or reputation block is not an interactive CAPTCHA. If the host can't operate a verification frame, Agent Verify preserves the resume point and uses the host's live-view or browser-takeover path.
Agent Verify is available under the MIT License.