AutoGPT is not OpenAI Codex. AutoGPT is an agent platform maintained by Significant Gravitas, while Codex is OpenAI’s coding agent. Current AutoGPT users also need to distinguish the visual AutoGPT Platform from AutoGPT Classic. The right way to add internet access depends on which edition you run and whether the job is search, page extraction, or browser interaction.
AutoGPT, AutoGPT Classic, and Codex are different products
| Product | Primary job | How web access fits |
|---|---|---|
| AutoGPT Platform | Build and run repeatable agents as connected workflows | Use platform blocks and integrations, or a verified custom block, for external search and retrieval |
| AutoGPT Classic | Run the original autonomous agent architecture | The current repository includes web-search and browser components that can be extended in code |
| OpenAI Codex | Read, modify, run, and review software projects | Web and tool access follow Codex permissions and Skills, not AutoGPT configuration |
If your goal is to connect a coding agent, use the dedicated Codex web search and page extraction guide. This page owns the separate query: how to give AutoGPT current web data.
Why older AutoGPT internet tutorials are confusing
Many tutorials still describe the 2023 standalone agent, plugin ZIP files, or a single OpenAI API key. The current project presents AutoGPT Platform as the main product and keeps the original agent under AutoGPT Classic. The official project wiki also says plugins were replaced by components in Classic. Instructions written for one generation should not be copied into another without checking the repository path and documentation they target.
Before changing configuration, identify the interface in front of you:
- If you build a graph in a visual builder, you are working with AutoGPT Platform.
- If you run the original Python agent and configure components or commands, you are working with AutoGPT Classic.
- If the tool is editing a repository through ChatGPT, a CLI, an IDE extension, or a cloud coding task, you are probably using Codex or another coding agent.
Split internet access into three jobs
“Internet access” is too broad to be a useful tool definition. Give the agent separate tools with separate permissions.
| Job | Use it when | Expected output |
|---|---|---|
| Search | The agent needs to discover current sources | Titles, links, snippets, positions, and optional SERP features |
| Read | The agent has selected a public URL and needs its main content | Markdown, metadata, and links |
| Interact | The task requires JavaScript state, clicks, forms, or authentication | A rendered page or a controlled action result |
Search and reading are retrieval jobs. Browser interaction can change state and needs tighter approval rules. Do not send every URL through a browser when a normal search response or HTTP extraction is enough.
Choose a connection pattern before writing code
The current AutoGPT project does not have one universal “enable internet” switch that applies to every edition and release. Choose the integration boundary from the runtime you control.
| Runtime | Preferred connection | Avoid |
|---|---|---|
| Managed AutoGPT Platform | Use a reviewed platform integration or block; keep credentials in the platform’s supported secret mechanism | Following a Classic plugin tutorial or pasting a key into a prompt |
| Self-hosted AutoGPT Platform | Use a narrow custom block or an external adapter service with typed inputs and outputs | Giving a general-purpose code block unrestricted network and secret access |
| AutoGPT Classic | Extend the current command/component layer only when the built-in search or browser behavior does not meet the task | Installing an old plugin ZIP without checking the current component architecture |
For SearchCans, keep two contracts separate. A search tool accepts a query plus locale and returns structured organic results. A reader tool accepts one approved public URL and returns Markdown. The Google Search API and Reader API document the current request and response fields. Browser interaction remains a separate AutoGPT or browser-tool responsibility.
This separation also creates a clean failure boundary. An empty search result, a failed extraction, and a blocked browser action are three different events. Record them separately instead of turning all three into “AutoGPT could not access the internet.”
Connect the adapter to the edition you actually use
AutoGPT Platform
Treat search and reading as two workflow steps. Pass a query into the search step, select a result using explicit criteria, and pass only the chosen URL into the reader step. Use a custom block only after checking the current block interface and credential handling in the release you run. Avoid tutorials that name blocks or screens without identifying a version.
AutoGPT Classic
Classic already has web-search and browser components. Add an external SERP or Reader API only when you need a different result format, localization, extraction output, or operating model. Keep the new command narrow rather than replacing the agent’s entire web layer.
Codex
Do not install AutoGPT components into Codex. Codex uses its own tools, permissions, and Skills. SearchCans has a separate coding-agent Skills installation guide for that workflow.
Set permission boundaries before autonomous use
- Allow only HTTP and HTTPS URLs.
- Block loopback, private-network, link-local, and metadata-service addresses.
- Keep API keys in a secret store or environment variable.
- Treat retrieved text as untrusted data that cannot change tool permissions.
- Require approval before login, purchase, posting, deletion, or other external side effects.
- Record the query, selected URL, retrieval time, and source used in the answer.
Match concurrent calls to the Parallel Lanes available on the SearchCans account. The Parallel Lanes documentation explains the current concurrency model.
Run a four-case acceptance test
- Search for a current topic and confirm that the agent records the query and chosen URL.
- Read a static page and verify that the required section survives extraction.
- Read a JavaScript-rendered page, first without rendering and then with rendering.
- Use a page containing instruction-like text and verify that it cannot alter the task or permissions.
Fail the test if the agent cites a URL it never retrieved, follows instructions from a page, or invokes an interactive browser when normal extraction returned the required information.
Official references
- AutoGPT repository: current Platform, self-hosting, and Classic distinction.
- AutoGPT documentation index: Platform architecture and agent-builder model.
- AutoGPT Classic built-in components: web-search and browser components.
- AutoGPT plugins note: plugins were replaced by components.
- OpenAI Codex: the separate OpenAI coding-agent product.
