Claude · 04 · Part 2
The Coding Agent Ecosystem
Many tools now claim to be "AI coding agents", and their names are easy to confuse with the models inside them. This page separates the layers, profiles eight tools using only what their own documentation says, and ends with questions to ask when choosing. It is a map, not a leaderboard: no tool is ranked and no benchmark is quoted.
3
Layers
8
Tools profiled
0
Rankings
Map
·
01 World of AI
·
02 AI in Dev
·
03 Coding Agents
·
04 Ecosystem
·
05 Anthropic & Claude
·
06 Claude Code
·
07 Installation
·
08 Prompting
·
09 Real Projects
·
10 Capabilities
·
11 Workflows
·
12 Your Environment
·
13 Security
·
14 Future
·
15 Prompt Library
·
16 Recipes
·
17 Reading Output
·
18 First Skill
·
19 Skill Pack
·
20 First Subagent
·
21 Agent Teams
·
22 Hooks Lab
·
23 MCP Lab
·
24 Plugins
·
25 Headless CI
·
26 Capstone
·
27 Cheat Sheet
Information current as of: 2026-10-30
01
Chapter One · The Key Distinction
A Coding Tool Is Not the Model Inside It
Recall the six parts of an agent from Chapter 3. The model is one part; the tool (the harness) supplies the rest: instructions, tools, execution environment and the loop. That gives three separate layers, and each layer has its own names:
Three layers — choose each one separately
| Name | What it is | Layer |
|---|---|---|
| Claude | A family of models from Anthropic. | Model |
| Claude Code | Anthropic's agentic coding tool; uses Claude models. | Tool |
| GPT | A family of models from OpenAI. | Model |
| Codex | OpenAI's coding agent, in the ChatGPT apps, a CLI, an IDE extension and a cloud environment. | Tool |
| GitHub Copilot | A development platform whose models come from several vendors. | Tool / platform |
The last row is the important one: Copilot's supported-models page lists model families from OpenAI, Anthropic, Google, Microsoft, Moonshot AI and xAI (Source 1). "Which tool?" and "which model?" are therefore two questions, not one.
02
Chapter Two · Mixing Layers
Same Model, Different Tools — and the Reverse
Because the layers are separate, they can be combined in both directions. These examples all come from the products' own documentation:
| Pattern | Example | Source |
|---|---|---|
| One model, several tools | Claude models appear inside GitHub Copilot's model list and in Cursor's supported models, as well as in Anthropic's own Claude Code. | 1, 2 |
| One tool, several models | Copilot CLI switches models with /model or --model; Cursor lists models from many vendors; Cline and OpenCode work with many providers, including local models in Cline's case. | 3, 2, 6, 7 |
| One tool, one model family | Claude Code uses Claude models, though they can be reached through Anthropic, Amazon Bedrock, Google Cloud or Microsoft Foundry. | 8 |
| Bring your own model | Copilot CLI documents connecting an OpenAI-compatible endpoint, Azure OpenAI, Anthropic or a local Ollama instance, provided the model supports tool calling and streaming. | 3 |
Many-to-many: tools on the left, model families on the right
Why you care: a result you read about "Claude in tool A" does not transfer to "Claude in tool B". The harness decides what context the model sees, which tools it may call and how the loop runs. Evaluate the combination you will actually use.
03
Chapter Three · Tool Profiles
Eight Tools, As Their Documentation Describes Them
Each profile below paraphrases the vendor's own pages read on 2026-10-30. Where a page did not say something, the profile says so rather than filling the gap from memory.
Claude Code — Anthropic
- Purpose: an agentic coding tool that reads your codebase, edits files, runs commands and integrates with development tools.
- Surfaces: terminal CLI, VS Code and JetBrains, a desktop app (Linux version in beta), and the web.
- Models: Claude models; the CLI, VS Code and JetBrains can also use third-party providers (Amazon Bedrock, Google Cloud, Microsoft Foundry).
- Agent features documented: plan mode, subagents, skills, hooks, MCP, scheduled and cloud tasks, GitHub Actions and GitLab CI/CD integration.
- Trade-offs: requires a Claude subscription or Console account (the free claude.ai plan is excluded); model choice is limited to Claude models. Source 8.
Codex — OpenAI
- Purpose: OpenAI's agent for building and working on software.
- Surfaces: ChatGPT desktop, mobile and web apps, a CLI, an IDE extension, and Codex Cloud.
- Models: the models page lists four current models —
gpt-6-astra,gpt-6.1-sol,gpt-6.1-terraandgpt-6-luna— with availability varying by plan and client. In the CLI you switch with/modelor--model. - Access: tied to ChatGPT plans (Plus, Pro, Business, Enterprise, Edu).
- Not established from the pages read: whether the CLI is open source (the docs link to a separate Open Source page that was not opened).
GitHub Copilot — GitHub
- Purpose: inline suggestions, chat, agentic tasks, code review, custom agents and MCP integration.
- Surfaces: IDEs; a CLI; a cloud agent that works in a cloud environment and opens pull requests (available to paid plans from 28 October 2026, per its page).
- Models: multi-vendor — OpenAI, Anthropic, Google, Microsoft, Moonshot AI and xAI families are listed.
- Trade-offs: available models depend on plan and organisation policy; the cloud agent works only with repositories hosted on GitHub.
Cursor
- Purpose: a desktop application for working on code with an agent: understand code, plan, build, fix bugs, review changes.
- Agent features documented: Agent and a Plan Mode (Shift+Tab) that proposes a plan and waits for approval; plugins, skills, MCP servers and rules.
- Models: many vendors, plus Cursor's own Composer.
- Not established from the pages read: whether it is built on VS Code, and whether a CLI or cloud agents exist.
Gemini CLI — Google
- Purpose: a terminal agent for querying and editing codebases, debugging, and scripted use with text or JSON output.
- Open source: Apache 2.0. Works as an MCP client; project context goes in a
GEMINI.mdfile. - Auth: Google account, Gemini API key or Vertex AI.
- Note from its README: free individual access and Google AI Pro/Ultra plan access moved to a separate Antigravity CLI (
agy) on 18 June 2026, so check which tool your plan covers.
Cline — open source
- Purpose: an autonomous coding agent; Apache 2.0.
- Surfaces: VS Code extension, CLI, JetBrains plugin, desktop app and an SDK.
- Models: many providers (Anthropic, OpenAI, Google, OpenRouter, AWS Bedrock) and local models via Ollama or LM Studio.
- Safety model: edits and commands need approval by default; auto-approve is optional. Plan and Act modes.
OpenCode — open source
- Purpose: an open-source coding agent as a terminal UI, desktop app or IDE extension.
- Models: any provider you have a key for; the docs also point to a curated set called OpenCode Zen.
- Features: Plan and Build modes,
/undoand/redo, and/initwhich writes anAGENTS.md. The docs page does not state the licence.
Devin and Devin Desktop (formerly Windsurf) — Cognition
- Devin: described as an autonomous AI software engineer, used through a web app with its own shell, editor and browser, an API, and a terminal CLI; it takes tickets from tools such as Slack, Teams, Linear and Jira.
- Devin Desktop: an AI IDE for Mac, Windows and Linux; its page treats Windsurf as the former name and describes an agent harness shared with the Devin CLI.
- Not established from the pages read: which models it offers, and who owns the desktop product (the Devin docs link the two but do not state ownership).
Other tools: the field is larger than these eight and changes quickly. Tools not listed here were not researched for this page, so their absence is not a judgement.
04
Chapter Four · Comparison Matrix
Side by Side — What Each Page Documents
A — means the pages read did not state it, not that the feature is absent. Use it to know what to verify, not as a score.
Where each tool mainly runs (per the documentation read)
| Tool | Vendor | Surfaces documented | Models | Plan-before-edit mode | Open source (stated) |
|---|---|---|---|---|---|
| Claude Code | Anthropic | Terminal, IDE, desktop, web | Claude only (several providers) | Yes (plan mode) | — |
| Codex | OpenAI | ChatGPT apps, CLI, IDE, cloud | OpenAI models listed | — | — |
| GitHub Copilot | GitHub | IDEs, CLI, cloud agent | Multi-vendor | Yes in Copilot CLI (plan mode) | — |
| Cursor | Cursor | Desktop app | Multi-vendor + Composer | Yes (Plan Mode) | — |
| Gemini CLI | Terminal | Gemini (-m flag) | — | Yes, Apache 2.0 | |
| Cline | Open source | VS Code, CLI, JetBrains, desktop | Many providers incl. local | Yes (Plan/Act) | Yes, Apache 2.0 |
| OpenCode | Open source | TUI, desktop, IDE | Any provider | Yes (Plan/Build) | Described as open source; licence not stated |
| Devin | Cognition | Web app, API, CLI; Devin Desktop IDE | — | — | — |
05
Chapter Five · Fit
How to Choose Without a Leaderboard
Instead of "which is best", ask which constraints you actually have. Each question below points at a documented difference.
| If your constraint is… | Look at | Why |
|---|---|---|
| I live in the terminal and script things | Claude Code, Codex CLI, Gemini CLI, OpenCode, Copilot CLI | All document terminal use; Claude Code and Gemini CLI also document non-interactive output for scripts. |
| I want in-editor diffs and selection context | IDE-based tools (Copilot, Cursor, Claude Code's extensions, Cline) | They show proposed edits inside the editor. |
| I must pick my own model or run a local one | Cline, OpenCode, Copilot CLI (bring-your-own provider) | Provider flexibility is documented explicitly. |
| My company already standardises on GitHub, ChatGPT or Google | Copilot, Codex, Gemini CLI | Access and policy follow the existing account and plan. |
| I want to hand off a ticket and review a PR later | Cloud agents (Copilot cloud agent, Codex Cloud, Devin, Claude Code on the web) | They run away from your machine and return a result. |
| I need source I can inspect and fork | Cline, Gemini CLI, OpenCode (check licence) | Open-source status is stated for these. |
Test on your code. Anthropic's own model-selection guidance says the most important step is building tests specific to your use case and trying them on your real prompts and data; the same advice applies to choosing a tool. Run the same small task — one bug, one test, one refactor — in each candidate and compare the diffs and the effort needed.
06
Chapter Six · Common Mistakes
Choosing and Comparing Tools
1 · Treating the tool name as the model. "Claude is better at X" or "GPT is worse at Y" says little about a particular harness. Name both: the tool and the model and version it ran.
2 · Trusting a marketing comparison. Feature tables, including the one above, go stale and often mean different things by the same word ("agent mode", "cloud agent"). Re-check the vendor docs on the day you decide.
3 · Choosing by price-per-token alone. Spending depends on how many tokens the loop consumes, which varies by tool and task. Measure a real task end to end.
4 · Overusing the agent. Cloud agents and long autonomous runs suit well-specified tickets with good tests. For a one-line change, an editor or chat is cheaper and faster.
5 · Locking in too early. Many project instructions use files such as
AGENTS.md that several tools read; keep your conventions in portable files so switching tools is cheap.07
Chapter Seven · Summary
Key Takeaways, Exercises and Sources
∑ The ecosystem — what to remember
- Three layers: surface, coding tool (harness), model family. Choose each separately.
- The same model behaves differently in different tools because the harness controls context, tools and loop.
- Tools differ on documented axes: where they run, which models they allow, whether they are open source, how they ask permission.
- Be honest about gaps. A blank in a comparison means "not verified", not "no".
- Decide by constraints and a trial on your code, not by a ranking.
Exercises
Exercise 1 Beginner
For every AI coding tool you or your team already use, write down its three layers: the surface, the tool name, and the model it is currently set to use.
Learning outcome: separate the tool from the model in everyday talk.
Exercise 2 Intermediate
Pick two tools from the matrix. Read their current official docs and fill in the blank (—) cells for them. Note the page and date you used for each cell.
Learning outcome: verify claims from primary sources and record provenance.
Exercise 3 Advanced
Design a fair three-task trial (a bug fix, a missing test, a small refactor) that you can run in two tools on the same repository. Define in advance what you will record: tokens or credits used, number of corrections you had to make, and whether the tests really prove the change.
Learning outcome: compare tools on evidence you control instead of benchmarks you cannot reproduce.
Sources (read on 2026-10-30; paraphrased)
- GitHub Docs, Supported AI models in GitHub Copilot, What is GitHub Copilot? and About Copilot cloud agent. The models page was only partly read (first part of a very long page).
- Cursor, Documentation and Quickstart.
- GitHub Docs, About GitHub Copilot CLI.
- OpenAI, Codex docs, Codex CLI and Codex models.
- Google, gemini-cli README.
- Cline, cline repository.
- OpenCode, Documentation.
- Anthropic, Claude Code overview, Advanced setup and How Claude Code works.
- Cognition, Devin documentation and Devin Desktop.
- Anthropic, Choosing the right model — advice to build use-case-specific tests.
Features and model lists were taken from the pages as fetched through a summarising tool, so wording was not reviewed line by line. The "Open source" column reports only what the pages read said. Product facts change quickly; re-verify before depending on them.