Three weeks ago I published Every AI Coding CLI in 2026: the complete map. It was a list for engineers in terminals. Most readers of AI coding tools are not engineers. They are builders, designers, and product people who want AI to help them ship without needing a terminal. This follow-up is the map for that audience.
If you are not a daily-terminal user but you want AI helping you ship (apps, websites, side projects, internal tools), the relevant question is not "which CLI". It is which surface fits how you actually work.
There is a clean three-surface taxonomy worth holding in your head:
| Surface | What it is | Where the work happens | Examples |
|---|---|---|---|
| Cloud Web | A browser tab; sync chat with code artifacts | Cloud, while you watch | claude.ai (Artifacts), ChatGPT Canvas, Gemini Canvas, Vibe |
| Desktop | An app on your machine; local files, local control | Local machine | Cursor, Devin Desktop, Zed, Continue, Claude Desktop |
| Cloud Agent | You hand off a task, walk away, check back | Cloud, async | Devin, Replit Agent, Bolt, Lovable, v0 |
The substantive divide is where the work happens, not whether the surface is GUI or terminal. Terminals are a desktop surface; the engineer's companion article covers the CLI subset inside Desktop for readers who live there. This piece is the rest of the map for everyone else.
The frame for this piece is builder versus engineer. Most articles in this space implicitly assume engineer. Builder means: you can read code, you have a clear product vision, you ship things, and you do not want to fight the tooling. Different surface, different shortlist.
What changed since the last article
Three weeks of churn worth knowing about.
- Windsurf is now Devin Desktop. Cognition acquired Windsurf for $250M in December; the rebrand landed June 2, 2026. The Cascade agent goes end-of-life July 1.
- Le Chat is now Vibe. Mistral renamed its consumer product May 28, 2026. Same model, same EU-jurisdiction story.
- Devin's price floor collapsed. Cloud agents used to be enterprise-only. Devin went from a $500/mo floor to $20/mo + $2.25 per ACU (about 15 minutes of agent work). This is the single biggest pricing event of the year for non-technical builders.
- Phind shut down January 16, 2026. They were the search-grounded coding tool. The stated reason ("commoditisation by foundation-model providers") is the same reason this category is worth re-mapping.
- Roo Code is archived (May 2026); migrate to Kilo Code.
- GitHub Copilot moved to usage-based AI Credits on June 1, 2026.
If you bookmarked anything from the old map, this is your patch.
Desktop IDEs: where most engineers actually live
The desktop tier has converged on a $20 entry price across almost every vendor. That is not a coincidence; it is the going rate for cloud-frontier-model use plus a wrapper.
| Tool | Price | What it is for |
|---|---|---|
| Cursor | $20 Pro, $60 Pro+, $200 Ultra | The most-used AI IDE. VS Code fork with agent + tab completion. Default for engineers who want a familiar IDE that just got smarter. |
| Devin Desktop (was Windsurf) | $20 Pro, $200 Max, $80+seat Teams | Post-acquisition consolidation. Replaces Cascade with Devin Local. Different bet from Cursor: Cursor optimises for the human at the keyboard, Devin Desktop dispatches more aggressively to cloud agents. |
| Zed | $10 Pro + $20 AI add-on | Rust-native, fast. Less feature-saturated than Cursor; that is the appeal. |
| Antigravity (desktop) | $20 Pro, $100 Ultra, $200 Max | Google's IDE play. Agent-first; bundles GPT-OSS 120B as a first-class option alongside Gemini 3.1 Pro and Claude 4.6. |
| Kiro | $20 Pro (1k credits), $40 Pro+ (2k) | AWS-aligned, spec-driven agentic engineering. If your stack is AWS this earns its keep. |
| Vibe (was Le Chat) | Per Mistral plans | The EU-sovereign desktop option. GDPR-native, runs on Mistral's own open weights. |
| Claude Desktop | Bundled with Claude paid plans | Anthropic's desktop app for Claude (macOS + Windows). Chat + code Artifacts + MCP connectors for files, repos, and tools. Not an IDE, but with MCP it sits closer to one than the simple-chat framing of last year. With Cowork enabled (next row), the same app becomes an autonomous agent host. |
| Claude Cowork | Included for paying Claude subscribers (macOS + Windows) | Desktop Agent. Lives inside Claude Desktop. You point Cowork at folders + connected tools (Google Drive, Gmail, DocuSign, FactSet, Zoom MCP, custom plugins), describe the outcome and cadence, and it completes multi-step tasks. Reads, edits, creates files. Went GA April 9, 2026 after a research preview at end of January. Enterprise tier has role-based access, group spend limits, OpenTelemetry, deployable plugins. |
For a builder picking right now, the live question is: do you want inline edits + tab completion (you are in the code, want it to keep up), agent dispatch on code (you describe a code task, walk away, come back), or knowledge-work agent (you point it at a folder + connected tools and it completes office-shaped work, not just code)?
- Inline + tab completion → Cursor or Zed.
- Agent-leaning, code-shaped work → Devin Desktop or Antigravity.
- Agent-leaning, knowledge-work shaped (analysis, reports, multi-tool flows) → Claude Cowork.
- Stack-specific tilt → Kiro for AWS.
- EU-sovereignty constraint → Vibe.
There is no single right answer. There is a right answer per project.
A taxonomy honesty note. Cowork breaks the clean Desktop / Cloud Agent split in the triad: it is a desktop app that behaves like an agent, with file system access and multi-step autonomy on the local machine. Devin Local (replacing Cascade in Devin Desktop) is moving in the same direction. The triad still helps as a first-cut sort, but the boundary between "Desktop" and "Cloud Agent" is dissolving as agent capability moves into local apps. Expect more of this. The surface where you sit and the agent layer that does the work are merging.
Desktop extensions: the OSS lane in your existing editor
If you already love VS Code or JetBrains and you do not want a new IDE, the extension tier is where open-source lives. All three of these support any provider you can point them at: Ollama for local, OpenAI-compatible for cloud, including the Chinese open-weights providers like DeepSeek and Qwen.
- Continue.dev: the leading OSS extension. Chat-first, not autonomous. Great if you want assistance, not a takeover.
- Cline: autonomous-agent class. The agent edits files, runs commands, asks for approval at each step. 5M+ installs.
- Kilo Code: fork of Cline + Roo (Roo is archived; Kilo is the live one to migrate to). 1.5M users, top OpenRouter consumer in 2026. Adds Orchestrator + Memory Bank.
If you are a builder who wants to keep your existing editor and just get smarter help, the right move is: Continue.dev for assistance, Cline or Kilo Code for autonomous tasks. Both flavours, free, open weights compatible.
Cloud Web: the canvas tier
This is the least-differentiated category in mid-2026. Every frontier lab's chat now has a Canvas-equivalent: a side panel where the model writes code while you watch, you can edit inline, both of you iterate. The differences are mostly model preference and ecosystem fit.
| Tool | Strength |
|---|---|
| claude.ai (Artifacts) | Best code artifacts in the browser. Strong for one-off scripts, learning examples, quick prototypes. |
| ChatGPT (Canvas) | Heaviest integration. Codex Cloud bridges from Canvas to the cloud agent. US free tier got ads in February 2026, a real signal if you care about data leakage. |
| Gemini (Canvas / Gems) | 1M+ context, deep Google Workspace integration. AI Ultra at $249.99/mo is the most expensive consumer tier here. |
| Vibe (web) | The EU-sovereign browser option. French jurisdiction, GDPR-native, runs on Mistral's open weights. Strongest if "no data leaves EU" is a hard constraint. |
| Perplexity | Search-grounded. Lighter coding, stronger research. |
| Grok (xAI) | Grok Build for code. Premium tier via X. |
Phind shut down in January 2026, and their stated reason ("commoditisation by foundation-model providers") is the honest read on the whole category. Every model vendor now has a Canvas-equivalent, so independent search-coding products had nowhere to stand.
For a builder, the practical advice: pick the lab whose model you already like, use that Canvas. Switching costs are zero; the differences are stylistic.
Cloud Agent: the category that woke up
This is the category that barely existed eighteen months ago and now has the most newcomers and the biggest pricing news. If you are a non-technical builder who can describe what you want in plain English, this is the surface that has changed the most for you.
| Tool | Price | What it is for |
|---|---|---|
| Devin | $20/mo + $2.25/ACU (~15 min) | The most autonomous; full sandboxed VM. Was $500/mo enterprise-only until April 2026. |
| Codex Cloud | Bundled with ChatGPT Plus / Pro | Async PR-from-prompt. 77.3% on Terminal-Bench 2.0. |
| GitHub Copilot Coding Agent | Bundled with Copilot from $10/mo | Cheapest cloud agent. Assign issue → PR. |
| Replit Agent | $20 Core, $95 Pro (10 parallel) | In-browser IDE + agent. Strong builder onboarding. |
| Bolt.new | From $20/mo (token-based) | Prompt-to-app via in-browser WebContainers. |
| Lovable | From $25/mo (message-based) | React + Supabase apps. Export to GitHub. |
| v0 | Bundled with Vercel paid | UI-component focus. Next.js native. |
| Devin Local | Bundled in Devin Desktop | Local-runtime cloud agent. Replacing Cascade July 1. |
| Antigravity Managed Agents | Bundled with Gemini API | Google's API-tier agent runtime. |
Devin's price floor collapse is the single most important pricing event for builders this year. A category that gated at $500/mo for the last 18 months now starts at $20 plus per-task billing. That makes autonomous-cloud-agent not a premium gate for the first time. Worth pausing on.
The prompt-to-app subcategory has stratified by stack:
- Lovable owns React + Supabase end-to-end.
- Bolt.new owns in-browser WebContainers (runs in your tab, no infra to set up).
- v0 owns Vercel + Next.js + the UI-component-first workflow.
If you are a builder shipping a customer-facing web app, those three are the live shortlist. Pick by which stack your future engineer is going to inherit when the app outgrows the prompt-to-app tier.
One honest constraint: none of the Cloud Agent tier is open-weights friendly today. These are all closed sandbox runtimes with proprietary models. If you want autonomous-agent capability and open weights, you are looking at Devin Local (alpha, partial) or building your own with one of the OSS CLIs from the engineer article.
A second axis worth tracking: where does the tool stop
A useful framing the May article missed is not just what tool, but where does it stop?
suggest → edit → commit → PR → deploy → live app
For a builder, the question maps differently than for an engineer:
- Cursor / Zed / Continue → suggest + edit. Human commits.
- Cline / Kilo Code → edit + commit. Human reviews.
- Cursor agent / Antigravity / Copilot Coding Agent → commit + PR. Human approves.
- Devin / Replit Agent → PR + sometimes deploy. Human checks the app worked.
- Bolt / Lovable / v0 → deploy + live app. Human iterates from the live result.
Pick by how far you want automation to terminate. Builders typically want it to terminate further down the chain than engineers do. That is exactly why the Cloud Agent tier matters more here.
Sovereignty: EU, USA, World
The political backdrop does not go away by ignoring it. For builders shipping to customers, regulated or not, here is the picture:
China-based providers dominate the cheap end of open weights: DeepSeek (V4 just shipped April 2026), Z.AI / Zhipu (GLM), Alibaba (Qwen), Moonshot (Kimi), MiniMax. Open-weights pricing dropped about 80% year-on-year, driven by this group. If you are pointing a desktop extension at a cloud open-weights model, the cheapest credible coding-capable model in the world right now is DeepSeek V4-Flash at $0.14 per million input tokens.
US-based hosters (Together, Fireworks, Groq, DeepInfra, OpenRouter) mostly do not own a model. They serve Chinese open weights at US-jurisdiction inference cost. Useful when you want Chinese-trained capability but US-or-EU data residency.
The single US-trained open frontier family is Meta's Llama 4 under the Community License (open, but not OSI-approved). There is a 700M monthly-active-user cap and a "no training competitors" clause that matters at scale. OpenAI's gpt-oss-120B and 20B (Apache 2.0) are the only fully-permissive US flagship-tier open releases.
EU-based. Mistral is the list. The only sovereign EU frontier-capable open-weights stack. GDPR-native. La Plateforme runs in EU data centres. If your compliance position is "no customer data leaves EU jurisdiction", the practical answer in mid-2026 is: self-host open weights on EU hardware, or use Vibe / Mistral. Short list.
This is unfortunate but not less true.
A note on the open-weights-first harness
Worth flagging because it shows up across both articles in this series: Ecodex is an Empirica-native CLI harness built around calibrated agent behaviour, with open-weights models as first-class citizens. It is alpha, opinionated, three GitHub stars at the time of writing, and not in the Desktop tier; it lives in the CLI category and is covered in the engineer article.
For builders, the relevant property is the compliance angle. The Ecodex team just shipped a compliance crosswalk mapping the substrate to EU AI Act, GDPR, and ISO 42001, framed as compliance as a computed property (the return value of compliance-report) rather than as a signed binder. If your project has a regulatory anchor that has to survive audit, that framing is worth a read.
The desktop side of this story is on the roadmap, not shipped. For now, Ecodex is a terminal thing.
What I would actually pick
Three reads for three builder contexts:
- Solo builder, web product, want speed. Bolt.new or Lovable for prompt-to-app; graduate to Cursor or Devin Desktop when the app gets real.
- Builder with an existing codebase. Continue.dev or Cline as a VS Code extension, pointed at whatever model your bill tolerates. DeepSeek V4-Flash for serious work, local Qwen3-Coder via Ollama for the no-bill option.
- Builder with a compliance constraint. Vibe in the browser, or self-hosted open weights on EU hardware via Cline + Ollama. Compliance-grounded path through Ecodex for the calibration story.
What I am not saying
- That the desktop tier is settled. Cursor is the default for a reason, but Devin Desktop is the most interesting bet on what comes after the inline-completion era.
- That Cloud Agents are ready to replace human engineers. They are ready to ship surface area (landing pages, internal tools, prototype apps). They are not ready to own a codebase.
- That open weights are at parity with closed frontier on every task. They are close on coding-specific benchmarks, behind on certain reasoning shapes, and the gap is narrowing fast.
- That sovereignty should come before capability. Pick by capability for the task; reach for sovereignty when the compliance position requires it.
The CLI side of this series is the companion piece, Coding CLIs in mid-2026: the engineer's map. If you are working primarily in a terminal, that is where to look next.
Comments and corrections welcome. Especially: tools I should have included, pricing I got wrong, builder workflows I missed.
Series: AI Coding Harness Map (2026). Pricing verified against official provider pages on 2026-06-21.
Top comments (0)