vibe-tools

A globally installed Node.js CLI by eastlondoner that hands a coding agent a set of other models to call — MIT, ~4.8k★ / 238 forks. Its own line for itself is “give AI agents an AI team and advanced skills.” Arrived here via a Dan Kornas X thread on 2026-08-04 (status 2084464575894770167); the claims below are checked against the repository, which is the page’s actual source.

What it does

vibe-tools install . walks through API-key setup, installs Chromium for browser automation, and writes or updates the instruction file for whichever coding environment you name — Cursor, Claude Code (claude-md), Codex, Windsurf, Cline, Roo. The capabilities it then exposes:

• Web search and deep research via Perplexity.
• Whole-codebase context via Gemini’s long window (the repo claims up to 2M tokens), used for repository analysis and search grounding.
• Browser operation via Stagehand — act, extract, observe — for driving and debugging a web app.
• Integrations for GitHub issues and PRs, Linear issues with context and comments, external dependency docs, YouTube analysis, and Xcode tooling.
• Provider coverage across OpenAI, Anthropic, Gemini, Perplexity, OpenRouter, ModelBox, xAI and Groq.

Why it matters here

It is a clean instance of a pattern this spoke keeps meeting from different directions: the harness is not the agent’s only model. claude-code and its peers own the loop, but vibe-tools sits beside them as a shell-callable roster the agent invokes — Perplexity for what the model doesn’t know, Gemini for what won’t fit in the window, Stagehand for what has to be clicked. Each is a capability the primary model lacks, bought by shelling out rather than by changing harness.

Two things about how it installs are the more interesting part. First, its integration surface is the instruction file — it configures .cursorrules, CLAUDE.md and their siblings rather than registering tools, which is the pre-MCP way of teaching an agent a capability and still works because instruction files are the one thing every harness reads. Compare agentic-awesome-skills arriving the same day from the opposite end: an MCP server the agent queries, built precisely because writing everything into the agent’s context stopped scaling. The two are the same problem — how does an agent learn it can do something — answered at opposite ends of the context budget.

Second, the whole-codebase-context feature is a bet that will age with model economics, not with this tool: it exists because the harness’s own model has a smaller window than Gemini’s. That is a gap vendors close on their own schedule.

Tier

T3. The arriving source was a promotional X thread (13 likes, ends in a newsletter pitch) that relays feature bullets; I re-derived the specifics from the repository, which is first-party and equally untested. Nothing here is measured: no evaluation of whether delegating research to Perplexity or context to Gemini improves an agent’s output, and the star count measures attention. The X thread’s own top comment — “these tools mostly fail after an hour of configuration” — is worth as much as the thread and is likewise unevidenced.

Related

dan-kornas · agentic-coding-harness · claude-code · claude-md · model-context-protocol · agentic-awesome-skills · browser-use · agent-scripts · synthesis