The Vibe Coding Landscape Changed Under Our Feet. Here Is the Map.

Two years ago, "AI coding tool" meant an autocomplete plugin and maybe a chat panel. Today it means a terminal agent that opens pull requests while you sleep, a spec-driven workflow with five slash commands, and a model marketplace where the open-weight option is sometimes the best one.

I maintain Awesome Vibe Coding, a curated list of AI-powered coding assistants, agents, models, benchmarks, and guides. This month I audited every single entry: every link, every price, every model name, every "free tier" claim. The results say a lot about how fast this space moves, so I want to share both the findings and the map that came out of it.

What an honest audit looks like

The list had about 240 entries and 246 links. Here is what checking all of them turned up:

• 25 dead links. Some were 404s on moved documentation. Some were parked domains for sale. One "Cursor for Beginners" tutorial link was, I regret to report, a Rick Astley video.
• 20 tools that no longer exist in the form the list described. Supermaven was sunset in November 2025 and folded into Cursor. Phind shut down in January 2026. Fig, Mentat, Height, Outerbase, and Ponicode are gone. Codeium became Windsurf, then became Devin Desktop under Cognition. Amazon CodeWhisperer became Q Developer, which is now being retired in favour of Kiro. Gitpod is Ona. Galileo AI is Google Stitch.
• Prices that were off by orders of magnitude. The OpenAI API line quoted a per-thousand-token price that, read as per-million, was a thousand times too high. GitHub Copilot was listed with an 8K context window. It now offers up to 1M.
• A comparison table where half the rows were about products that had changed hands.

None of this is a criticism of anyone who contributed. It is just what happens when a field reinvents itself every six months and a list does not. If you maintain an awesome list in this space, budget for a full re-verification at least twice a year, and check GitHub's API for archived repositories, not just for HTTP 200s.

The new centre of gravity: terminal agents

The biggest structural shift is that the most-used AI coding tools of 2026 are not editor plugins. They are agents you run in a terminal, and every major model vendor now ships one:

• Claude Code from Anthropic, with subagents, hooks, skills, and MCP support.
• Codex CLI from OpenAI, open source under Apache 2.0, with a cloud mode for background tasks.
• Gemini CLI from Google, also Apache 2.0, with a generous free tier.
• GitHub Copilot CLI, with native repository, issue, and PR access.
• Muse Code from Meta, Grok Build from xAI, Qwen Code from Alibaba, Kimi Code from Moonshot, and Mistral Vibe from Mistral.

Alongside the vendor tools, a set of open-source, provider-agnostic agents has grown up fast. OpenCode is MIT-licensed and works with any model, including local ones. DeepSeek Harness treats providers, tools, and context injection as swappable plugins. Pi is a deliberately minimal TypeScript agent that writes its own extensions. Crush from Charm replaced their earlier mods tool with a full agent.

The practical upshot: if you have API keys or a subscription to any frontier model, you can have a capable coding agent in your terminal in under five minutes, and you can usually swap the model without changing your workflow.

Spec-driven development is the new prompt engineering

"Vibe coding" started as a joke about accepting whatever the model produced. The tools that have taken off in 2026 are the ones that add structure back:

• GitHub Spec Kit gives you constitution, specify, plan, tasks, and implement commands.
• OpenSpec keeps change proposals and specs next to the code.
• BMAD-METHOD assigns agent roles across the whole lifecycle.
• Superpowers packages brainstorming, planning, test-driven development, and review as reusable skills for Claude Code, Cursor, Codex, and Copilot CLI.

These are not competing IDEs. They are methodology layered on top of whichever agent you already use. That is a healthier place for the ecosystem to be than a dozen incompatible editors.

MCP tooling has become its own category

The Model Context Protocol went from a spec to infrastructure. The list now has a proper set of MCP tools:

• Context7 injects current, version-specific library documentation into agent prompts, which fixes a whole class of "the model trained on the old API" bugs.
• Serena gives agents symbol-level code navigation and refactoring through language servers.
• FastMCP is the standard way to build a server in Python.
• MCP Inspector is the reference debugging tool.

Open-weight models are a real option now

The model section needed the most rewriting. The 2024 entries were CodeLlama, StarCoder 2, WizardCoder, and CodeGen. All of those are legacy now. The current open-weight lineup for coding is Qwen3-Coder, DeepSeek V4, Devstral 2 from Mistral, gpt-oss from OpenAI, Kimi K2, GLM-5, and Phi-4. Several of them run on a single workstation through Ollama or LM Studio, which means "offline" is a genuine checkbox in the comparison table again, not a theoretical one.

Benchmarks moved too

HumanEval is saturated. The benchmarks people actually cite in 2026 are SWE-bench Verified and its harder successor SWE-bench Pro, Terminal-Bench for agentic terminal work, the Aider Polyglot suite, and SWE-rebench, which refreshes monthly to resist contamination. Chatbot Arena became LMArena and then simply Arena, with a dedicated code and web-development leaderboard. The list links to all of them.

Free tiers that are actually free

Because "free" was the most misreported claim in the old list, the refreshed version only lists free tiers verified against vendor pricing pages. As of September 2026 they include GitHub Copilot Free, Cline, OpenCode, Gemini CLI, the Windsurf plugin's unlimited autocomplete, Zed with your own keys or local models, Tabby's self-hosted community edition, Bolt.new's monthly allowance, and the JetBrains AI tier bundled with their IDEs.

How the list stays current

The repository has three GitHub Actions: a daily job that checks links, recounts tools, and opens a pull request; a weekly job that searches GitHub for trending new AI coding repos and files an issue; and a lint job that enforces the awesome-list format. Part of this month's work was fixing all three, since a mismatched action tag and a missing dependency had quietly broken them.

Contributions are welcome, and the bar is simple: the tool must actually exist, must be maintained, must have documentation, and the description must be objective rather than a pasted marketing tagline. Pull requests that disclose vendor affiliation are fine. Pull requests that describe a source-available licence as open source get a polite request for a reword.

Where to start

If you are new to all of this, the shortest path I can suggest is:

1. Pick one terminal agent. Claude Code, Codex CLI, or Gemini CLI if you want a vendor's tool; OpenCode if you want to own the model choice.
2. Add one methodology layer. Spec Kit or Superpowers will do.
3. Add Context7 so the agent stops guessing library versions.
4. Run your first real task against a benchmark-style problem in your own repo, not a toy.

The full map, with sections for IDE integrations, agents, browser automation, code review, testing, DevOps, databases, learning resources, benchmarks, model providers, and enterprise tooling, is at github.com/tysoncung/awesome-vibe-coding. Stars help other people find it. Pull requests keep it honest.