Skip to content

Decision tree: which surface?

Last verified: 2026-05-06 · Drift risk: medium

There are four broad categories of surface for building with AI agents: chat apps (subscription-tier), custom assistants built on top of those apps, desktop/CLI clients, and the raw API/SDK. This page helps you pick the right starting point.

The core questions

Work through these questions in order. Stop at the first recommendation that fits.

1. Do you just need a better chat experience today, with no custom tooling?

Yes → Start with a chat app.

  • Use ChatGPT Projects to organize conversations by topic, share custom instructions, and keep files in scope across sessions.
  • Use Claude.ai for long-context work and conversational analysis.
  • Use Gemini app + Google AI Studio when you need deep integration with Google Workspace data.

No → Continue to question 2.

2. Do you have a coding workflow — editing code, running tests, refactoring a repo?

Yes → Use a coding-focused agent surface.

  • Claude Code runs in your terminal and has direct filesystem and shell access. Best if you work primarily in Claude.
  • OpenAI Codex CLI is a terminal-native coding agent for OpenAI models.
  • Gemini CLI provides a similar terminal experience for Gemini models.
  • GitHub Copilot's coding agent operates inside GitHub, handling issues and pull requests without leaving the GitHub UI.

No → Continue to question 3.

3. Do you need to connect external tools (databases, APIs, file systems, SaaS products)?

Yes → You need MCP or a connector layer.

  • With a desktop client: Claude Desktop + one or more MCP servers is the most common local MCP setup. Anthropic also provides connectors for common SaaS tools.
  • With a CLI: Gemini CLI, Claude Code, and Codex CLI all support MCP servers via configuration files.
  • With a custom assistant: Custom GPTs can call external APIs via Actions. Claude.ai Projects work with connectors.

No → Continue to question 4.

4. Do you need long-running tasks, custom logic, or a multi-agent pipeline?

Yes → You need the API or an SDK.

  • OpenAI Agents SDK provides primitives for agents, tools, handoffs between agents, and tracing. Start here for Python-based multi-agent systems on OpenAI models.
  • Anthropic's API supports tool use and multi-turn conversations natively. Combine with the MCP spec for standardized tool definitions.
  • Google AI Studio provides access to Gemini models with function calling for code-based workflows.

No → You are likely in "better chat" territory. Go back to question 1.

5. Do you need everything to run locally (data residency, air-gap, or privacy)?

Yes → Read the Local-first path. Use stdio MCP servers, CLI tools, and local file access. Note that the model itself typically still runs in the cloud unless you run a self-hosted model.

6. Are you building for a team or org?

Yes → Read the Team path. This changes the surface choice: you need SSO, shared resources, audit logs, and key management that personal subscriptions do not provide.

Summary table

Your situation Recommended starting point
Better chat, no tooling ChatGPT Projects or Claude.ai
Coding and terminal work Claude Code, Codex CLI, or Gemini CLI
Connect external tools, no code Claude Desktop + MCP or Claude connectors
Connect external tools, some code Custom GPT Actions or Claude Projects + connectors
Custom pipeline, Python OpenAI Agents SDK or Anthropic API
Everything local CLI tools + stdio MCP servers
Team or org Enterprise tier on whichever platform fits above

What to do next