--- title: "The Last Twins Left in the Terminal — The Logic of People Who Stay With Aider" date: 2026-10-10 model: qwen3.6-plus author_type: human category: reviews summary: Aider has no GUI and no autocomplete. In exchange it scores "Value 9.0" as the most cost-rational AI coder, on git-native, model-agnostic, and token transparency. Here it is with its limits: the 25k token context cliff, manual REPL driving, and no MCP. tags: Aider,terminal coding,git native,model agnostic,BYOK,REPL,user reviews,real user reviews,AI agent --- # The Last Twins Left in the Terminal — The Logic of People Who Stay With Aider One review score is the starting point. "Ease of use6.0, Features8.0, **Value9.0**." (Litmus,2026) This tool's usability is low and its value is top-tier. That contradiction is all of Aider. The reasons left by users who "switched to Cursor or Cline and then quietly came back" are similar: "transparency of the loop and cost discipline." This post evaluates Aider only as a tool for people who live in the terminal, and sorts out what you give and what you get within that. --- ###1. What you get **git-native.** Every AI edit auto-commits (or selectively). A month later you have history you can blame. "Cursor and Cline make you review and commit by hand, so the temptation to bundle20 unrelated changes into one commit is real." **Model-agnostic.** `/model` swaps Claude, DeepSeek, or GPT mid-session. BYOK is the default, so no subscription lock-in. **Cost transparency.** Token usage printed every turn. Measured records of $2–4 a day sit in several reviews. **Terminal survival.** Works beside SSH, Docker, Helix, and Vim. A different tier from tools that need a GUI process. "Like a tricycle — very good, once it's adjusted." **repo map + lint/test.** The tree-sitter-based repo map keeps the model from pasting whole files. After a change, lint and test run by default and it asks the model to fix things. --- ###2. What you don't get **No autocomplete.** No inline ghost text. "If half your usage is finishing lines with Tab like Copilot, Aider can't replace that." **No agent loop.** Hand it "fix this tricky test" and **you have to drive it in the REPL** — run the test, check the output, revise, repeat. You can get close with `--yes-always`, but it isn't Claude Code's autonomy. The user community's counter, though: "An explicit loop is actually safer. Cursor agents give up halfway and gut half your codebase; Aider has no autonomy, so it can't give up." **Context is manual, with a 25k cliff.** You `/add` files directly, and records repeatedly note reliability drops past 25,000 tokens. Context management is a skill, not a hassle. **No MCP.** "The biggest downside is no MCP" is a repeated opinion. It sits outside the connector ecosystem — Playwright, Context7, and the rest. **No GUI diff.** You read a terminal diff or go back to git. It asks for the habit of trusting fast-scrolling output and reviewing later. --- ###3. Who stays and who leaves The classification the reviews agree on. | Stays | Leaves | | --- | --- | | Already lives in the terminal | Depends on GUI and tab autocomplete | | Clean git history matters | Wants the agent to finish it for them | | Codebase too big for IDE indexing | Onboarding non-technical colleagues | | Needs air-gapped or local-model coding | MCP ecosystem is essential | "The activation energy (API key, slash commands, conventions file) is higher than alternatives" is the core analysis. Aider loses new users and keeps existing terminal users — that's the structure. --- ###4. Conclusion — the most rational cost, the steepest entrance Aider's user record ends in two sentences. **For someone for whom the terminal and git feel natural, it's the most cost-rational AI coder choice there's been.** Value9.0 isn't exaggeration. **For everyone else, it's the wall of a tool with no autocomplete, no MCP, and no GUI diff.** The25k context cliff and the manual REPL driving are structure, not style. Aider is often used not as a replacement but as a **side tool.** Cursor takes tab and point edits, Claude Code takes heavy autonomous work, Aider takes SSH, auto-commit, and model switching. The three don't overlap and divide up the shares. Related posts: [Cline user reviews](/reviews/2026-10-10-cline-user-reviews/) · [OpenCode user reviews](/reviews/2026-10-10-opencode-user-reviews/) · [Aider agent hub entry](/agents/aider/) --- ### Sources - Overall review: [Aider Review (Litmus,2026-06)](https://litmustools.com/review/aider/) · [Aider Review (Awesome Agents,2026-02)](https://awesomeagents.ai/reviews/review-aider/) - Three months of real use: [Aider review2025 (TopInsight)](https://topinsight.co/ai-coding/aider-review-2025/) - Comparison and MCP: [Aider Review2026 (AICoderScope,2026-05)](https://aicoderscope.com/blog/aider-review-2026/) · [I Used Aider for a Week (aimadetools,2026-05)](https://www.aimadetools.com/blog/aider-one-week-review/)