Command-line and agent tools
Choose the appropriate AGTCloud development tool based on usage scenarios, protocols, and permission scopes.
AGTCloud integrates with command-line agents, IDE extensions, desktop clients, and configuration management tools. When choosing a tool, first confirm its workflow and protocol, then evaluate which files, commands, and external services it can access.
This page helps you choose a guide without repeating each tool's installation and configuration steps. In the selected guide, follow its Base URL, authentication, and model requirements to complete the integration.
Agent tools may perform local operations
Some tools can read and modify files, run shell commands, or access the network. For first-time access, use a test directory that does not contain sensitive data, disable auto-approval, and start validation from read-only tasks.
Select by tool
| Tool | Form | AGTCloud protocol | Best suited for |
|---|---|---|---|
| Claude Code | Terminal coding agent | Anthropic Messages | Existing Claude Code workflows that use Messages and Content Blocks |
| Codex | CLI and desktop app | OpenAI Responses | Responses, Typed Items, or Codex project workflows |
| Gemini CLI | Terminal agent | Gemini Generate Content | Applications that use Gemini contents and parts |
| Cline | IDE extension | OpenAI Chat Completions | Reviewing agent operations in VS Code-compatible editors or JetBrains IDEs |
| Cherry Studio | Desktop chat client | OpenAI Chat Completions | Graphical multi-model chat and basic connection verification |
| OpenClaw | Self-hosted Gateway, TUI, and Control UI | OpenAI Chat Completions | A local gateway and extensible personal-agent environment |
| CC-Switch | Local configuration manager | Anthropic Messages, OpenAI Responses | Saving and switching provider configurations for Claude Code and Codex |
| Hermes Agent | Terminal agent | OpenAI Chat Completions | Custom endpoints, terminal sessions, and tool-call workflows |
Protocol names do not represent the model vendor. The same model may provide multiple protocol entry points, but tool calls, multimodality, inference, and context capabilities are still subject to the actual support of the current model.
Select by task
| Goal | Recommended guide | Why |
|---|---|---|
| Perform code analysis and modification in the terminal | Claude Code, Codex, Gemini CLI | All three have complete terminal workflows, but use different protocol structures |
| Gradually review documents and command operations in the IDE | Cline | Operation plans, file changes, and authorization processes are all centralized within the editor |
| Use a graphical interface for regular model conversations | Cherry Studio | After configuring the service provider and model, the connection can be verified via the chat interface |
| Run local, self-hosted agents | OpenClaw, Hermes Agent | Providing Gateway or CLI agent operation requires stricter host permission control |
| Manage multiple Claude Code or Codex configurations | CC-Switch | Suitable for saving and switching providers locally, without replacing the client that actually initiates the request |
| Connecting external tools and data sources | MCP Access and Security | MCP is a protocol between the client and the tool service, not equivalent to a model API |
Confirm before connecting
1. Confirm the usage form
Decide whether you need a terminal, IDE, desktop chat, or self-hosted gateway. Don't choose based solely on the tool name; first confirm whether it can adapt to your team's approval and permission management methods.
2. Confirm the protocol
Copy the model ID available for the current API Key from the Model Marketplace and confirm that the model supports the Chat Completions, Responses, Anthropic Messages, or Gemini Generate Content entry required by the tutorial.
3. Prepare separate credentials
Create an API Key on the API Keys page that is easy to identify and deactivate. Do not write the real Key into the repository, screenshots, logs, or shared configurations.
4. Review permissions
Verify the directories, commands, networks, and external services accessible by the tool. By default, auto-approve is disabled, and only the minimum range needed to complete the current task is enabled.
Phased validation
| Stage | Suggested task | Pass criteria |
|---|---|---|
| Basic connection | Only reply with a fixed text without calling the tool | It can return content, and the Provider, Base URL, and model meet expectations |
| Read-only files | Read and summarize README.md in the test directory | No files are written and no unapproved commands are executed |
| Controlled tools | Perform a revocable operation with a clear scope of impact | The client displays the operation content and confirms the results, which match expectations |
| Project Tasks | Handle real projects under version control protection | Changes can be reviewed, rolled back, and can pass existing project checks |
Do not assume that "client successfully started" means access is complete. At the very least, confirm that the actual request uses the expected provider, model, and protocol, and check that the tool does not exceed the authorization scope.
Common selection issues
| Problem | Handling recommendations |
|---|---|
| Chat works, but tools do not | First confirm that the target model and protocol support tool calls, then check the client version and tool permissions |
| The same model performs differently in different tools | Compare protocols, system prompts, tool schemas, context, and client default parameters |
| Configured correctly but still requesting other Providers | Fully restart the client and check the user-level configuration, environment variables, and the currently enabled Provider |
Base URL returns 404 | Use the guide to confirm whether the client expects the API Origin or the /v1 Base URL, and whether it appends the operation path automatically |
| You are unsure whether to enable automatic approval | Keep it disabled. Enable approvals selectively only after the task, directory, and command boundaries are clearly defined |
| Requires connection to databases, file systems, or business services | Read MCP Access and Security first, and verify the source and permissions of the service |
Safety advice
- Download the installation package or command from the tool's official channel and read the version instructions before upgrading;
- Use separate API Keys for different tools so each can be revoked and audited independently;
- Running tasks that modify files under version control or backup protection;
- Do not submit passwords, customer information, production data, or other unnecessary sensitive information to the model;
- Check whether tool logs, session exports, and crash reports contain Key, path, or business data;
- After stopping the tool, delete the credentials saved within it and revoke any permissions you no longer need.
If you have already identified the tool, please start with CC-Switch Integration Tutorial, or directly select the corresponding tool from the table above. Please refer to API Reference for interface fields and response structure.