ChatGPT agent is the OpenAI feature that lets ChatGPT carry out multi-step tasks on the web for a user. It was announced on July 17, 2025 and builds on Operator, the earlier browsing agent, and deep research. Operator is listed here as a synonym. OpenAI said the Operator research preview would be sunset after a few weeks, and it now marks the launch post as outdated (as of October 2026).
At a glance
- Operator
- OpenAI
- Purpose
- multi-step tasks on the web on behalf of a user
- robots.txt token
- not listed on OpenAI's bots page
- Respects robots.txt
- not listed on OpenAI's bots page
Last checked:
What can ChatGPT agent do?
OpenAI describes a set of tools in one agent: a visual browser that works through a page’s interface, a text-based browser for simpler retrieval, a terminal and API access. The agent can fill in forms, work with spreadsheets and use connected services.
For site owners this means visits that can look like a person using a browser. They originate from a single user request and follow no crawl schedule.
| Bot | Operator | Purpose | robots.txt token | robots.txt according to the operator | IP list published? |
|---|---|---|---|---|---|
| Google-Extended | Controls Gemini training and grounding (no crawler of its own) | Google-Extended | robots.txt token only; fetching is done by existing Google crawlers | n/a (no user agent of its own) | |
| GPTBot | OpenAI | Training | GPTBot | Disallowing it signals content should not be used to train foundation models | yes (openai.) |
| OAI-SearchBot | OpenAI | Search (ChatGPT) | OAI-SearchBot | opted-out sites are not shown in ChatGPT search answers | yes (openai.) |
| ChatGPT-User | OpenAI | User-triggered fetch | ChatGPT-User | robots.txt rules may not apply, says OpenAI (user-initiated) | yes (openai.) |
| ClaudeBot | Anthropic | Training | ClaudeBot | yes, Anthropic says its bots honor robots.txt | yes (claude.) |
| Claude-SearchBot | Anthropic | Search (Claude) | Claude-SearchBot | yes, Anthropic says its bots honor robots.txt | yes (claude.) |
| PerplexityBot | Perplexity | Search (Perplexity), not used for model training | PerplexityBot | controllable via robots.txt; the docs contain no explicit compliance statement | yes (perplexity.) |
| Applebot-Extended | Apple | Controls AI training (does not crawl, says Apple) | Applebot-Extended | robots.txt token only, applies to content Applebot already crawled | n/a (no crawler of its own) |
| CCBot | Common Crawl | Open web crawl archive | CCBot | can be blocked via robots.txt, says Common Crawl | yes (index.) |
| Meta-ExternalAgent | Meta | Training foundation AI models, improving products | meta-externalagent | can be blocked via robots.txt, says Meta | not stated by the operator |
A selection of ten AI bots from major operators, not a complete list. All values come from each operator's own documentation (as of 8 October 2026). The table does not show whether a bot actually complies with robots.txt; it only reflects what operators state. User agents can be spoofed; published IP lists or reverse DNS identify genuine requests. Sources: OpenAI: Overview of OpenAI Crawlers; Anthropic: Does Anthropic crawl data from the web, and how can site owners block the crawler? (updated 2026-04-07); Perplexity: Perplexity Crawlers; Google: Google's common crawlers (last updated 2026-07-14); Google: Overview of Google crawlers and fetchers (last updated 2026-06-12); Google: Verifying Google crawlers and fetchers (IP range files); Apple: About Applebot (2026-09-04); Meta: Meta Web Crawlers; Common Crawl: CCBot. Last checked: .
How does it show up on a website?
According to OpenAI’s help center, the agent’s Cloud browser signs outbound requests with Web Bot Auth, meaning HTTP Message Signatures (RFC 9421). Site operators can verify the signature instead of relying on the user agent string. OpenAI’s bots page lists GPTBot, OAI-SearchBot, OAI-AdsBot and ChatGPT-User with their user agents and IP lists. ChatGPT agent has no entry of its own there. For ChatGPT-User, which fetches pages on user actions, OpenAI writes that robots.txt rules may not apply. The documentation gives no statement on whether ChatGPT agent follows robots.txt, so no rule can be named here as a reliable control.
Example
A person asks ChatGPT agent to compare three office furniture suppliers. The agent opens each supplier’s pricing page in a browser, reads prices and delivery times and puts them in a table. In the log files this appears as short sequences of page requests across several domains.
Common mistakes
- Stripping the signature headers. According to OpenAI, proxies between client and server must pass on the
Signature,Signature-InputandSignature-Agentheaders. Otherwise the request cannot be verified and may be blocked. - Allowing without verifying the signature. OpenAI advises allowing requests only after the signature has been verified successfully.
FAQ
Is ChatGPT agent a crawler? It does not scan the web on its own. It acts on instructions from one person.
How is it different from ChatGPT-User? ChatGPT-User fetches pages to support an answer. The agent also performs actions, such as filling in forms.
Sources
- Introducing ChatGPT agent – OpenAIopenai.com
- Introduction to ChatGPT Agent – OpenAI Developer Communitycommunity.openai.com
- OpenAI Crawlers and fetchers – OpenAI Developersdevelopers.openai.com
- ChatGPT agent – OpenAI Help Centerhelp.openai.com
- ChatGPT agent allowlisting – OpenAI Help Centerhelp.openai.com
Last updated: