核心 · Key Idea
In one line: In a year, coding agents went from "Tab completion" to "reading repos + editing many files + running tests". Cursor / Windsurf are IDEs; Cline / Continue / Aider are VSCode plugins + CLIs; Devin / OpenAI Codex are cloud agents.
Cheatsheet of mainstream tools#
- Cursor / Windsurf
- VSCode forks — **full-featured IDE**: tab completion, Chat, multi-file Agent, Terminal Agent. Subscription.
- GitHub Copilot
- VSCode/JetBrains plugins + cloud IDE. Completion + Chat + Workspace Agent.
- Cline / Roo Code
- VSCode plugins with built-in agent loop (read → edit → run → observe). BYO API.
- Continue
- Open-source VSCode/JetBrains plugin; freely configurable backend.
- Aider
- CLI + git-diff style — **fits the git workflow**.
- Devin / Manus / Genspark
- Cloud-native true agents with browser + shell + IDE; can run hours-long tasks.
- OpenAI Codex CLI / Claude Code
- Local CLI agents.
- JetBrains AI / Junie
- JetBrains' first-party agent across the JetBrains family.
Analogy#
打个比方 · Analogy
Tab completion = type half a sentence, it finishes;
Chat-in-IDE = an intern next to you answering questions;
Agent-in-IDE = the intern reads the project, edits files, runs tests, opens a PR;
Cloud agent = you hand them the task and they work overnight on their own machine.
Key capability comparison#
Repo Map / IndexingRepo map
Embed + structure the whole repo so the model can 'jump to definition'. Cursor / Cline both have this.
Tool UseTool use
shell / file-edit / browser / git. Determines how much an agent can actually do.
Diff-based EditDiff editing
Aider / Cline use diffs instead of full-file overwrite — **token-efficient + easy to revert**.
Context CompressionContext compression
For long tasks: keep relevant context via summaries / retrieval.
BYO ModelBring-your-own model
Cline / Continue / Aider freely swap OpenAI / Claude / DeepSeek / local.
MCPModel Context Protocol
Anthropic's standard for LLM-tool integration. Cursor / Cline / Claude Desktop all support it.
Choosing#
Practical notes#
- Rule one: give agents tasks with clear verification (lint / tests / build pass = done) to reduce hallucination.
- Small tasks → local IDE agent. Cline + DeepSeek-V3 / Claude / GPT-4o handles routine code.
- Large / long tasks → cloud agent. Devin / Manus run in their own environments — parallel + persistent beats local.
- Wire MCP servers: filesystem / git / web-search / DB query — tools cap agent capability.
- Code review is the floor. All agent edits go through PR; never push to main directly.
- Cost control: cheap model for tab completion, strong model for big changes — tier within one project.
- Privacy / compliance: private repos → "no-train / data-residency" services, or self-host + local / private API.
Easy confusions#
Tab completion (early Copilot)
Single-point inline continuation.
Reactive to your current cursor.
Reactive to your current cursor.
Agent-in-IDE
Actively reads repo + edits many files + runs commands.
"Hand it the entire task."
"Hand it the entire task."