Updated
Vibe coding

Context & Skills for Coding Agents

Give your coding agent what it can't see on its own — current library docs, a map of the codebase, a working method, a real browser — through MCP servers, skills and packed context.

Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Trends

Context7 gained the most ground in package downloads: +86% against the median npm package over a year.

Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.

ToolDownloadsStars
Playwright92.2M/wk+17% vs npm★ 97.5k
Chrome DevTools MCP1.8M/wk★ 53.8k
Context7704.1k/wk+86% vs npm★ 63.3k
Repomix90.6k/wk−7% vs npm★ 28.9k
Serena34.4k/wk★ 30.4k
BMAD Method18.1k/wk−18% vs npm★ 54.1k
Beads6.6k/wk★ 28k
Superpowers—★ 301.1k

The real choice

What the agent is missing. Docs servers stop it writing code against an outdated version of a library; code-understanding tools help it find its way around a large or unfamiliar repo; skills and methods change how it works (spec first, tests first) rather than what it knows; browser servers let it open the app it just built and check the result. Each one you add costs context and tokens on every task, so pick for the gap you keep hitting.

Pick by situation

Tap the ones that are you — the tools that fit light up below.
  • The agent keeps using APIs that changed in a newer version of a libraryContext7
  • You're working in a large or unfamiliar codebase and the agent keeps missing related codeSerenaDeepWiki
  • You want to paste a whole repository into a chat model or a model without file toolsRepomix
  • The agent jumps straight to code and you want it to plan and test firstSuperpowers
  • The agent should open the app it built, click through it and check for errorsPlaywrightChrome DevTools MCP

Compare the tools

Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.

ToolBest for
Library docs1
Context7service · free tierCurrent, version-specific documentation for thousands of libraries, fetched by the agent through an MCP server, CLI or API so it writes code against the docs instead of its training data.14 open source704.1k/wk+86% vs npm

Use it whenYour stack moves faster than the model's training cutoff and the agent keeps reaching for old APIs.

Trade-offThe free plan covers 1,000 calls a month on public libraries; private repositories and more calls are paid, and self-hosting is Enterprise only.

No maker's product we track shows it for this yet · in 14 open-source projects.

Code understanding2
Serenalibrary · open source · free tierIDE-style tools for the agent, backed by language servers — find a symbol, its references, edit or refactor at the symbol level — instead of grepping and rewriting whole files.8 open source34.4k/wk

Use it whenYour codebase is large enough that text search wastes the agent's context, in a language with a good language server.

Trade-offOpen source (GPL-3.0) and runs locally, but you install a Python tool and language servers yourself, and agents with strong built-in code search gain less.

No maker's product we track shows it for this yet · in 8 open-source projects.

DeepWikiservice · free tierAuto-generated wikis of GitHub repositories, with architecture diagrams and links to source, that you or your agent query through chat or its MCP server.3 open source

Use it whenYou need to understand a library or open-source dependency quickly, or give the agent that understanding.

Trade-offFree only for public repositories; private repositories need a Devin plan, and the wikis are generated by Cognition's service, not by you.

No maker's product we track shows it for this yet · in 3 open-source projects.

Packed context1
Repomixlibrary · open source · free tierPacking a repository into one XML, Markdown or JSON file, with token counts and likely secrets left out, to paste into a chat model.5 open source90.6k/wk−7% vs npm

Use it whenYou want a model without file access — a chat window, a long-context API call — to see the whole codebase at once.

Trade-offA snapshot, not live access; large repositories quickly exceed a model's context, so you filter what goes in.

No maker's product we track shows it for this yet · in 5 open-source projects.

Skills and methods1
Superpowerslibrary · open source · free tierAn open-source skills library that has the agent draft a spec with you, write a plan, then implement it test-first with subagents.No evidence yet★ 301.1k

Use it whenThe agent jumps straight to code on tasks that needed a plan, and you use Claude Code or another agent with skills support.

Trade-offSlower and more token-hungry than a plain prompt on small changes; GitHub is its only home.

No maker's product we track shows it for this yet.

Browser for the agent2
Chrome DevTools MCPlibrary · open source · free tierGoogle's MCP server and CLI that gives the agent a live Chrome with DevTools — performance traces, network requests, console messages, screenshots.26 open source1.8M/wk

Use it whenYou want the agent to debug what the browser sees — a slow page, a failing request, a console error — not just click through flows.

Trade-offChrome only; it sends usage statistics to Google unless you pass the opt-out flag, and it exposes the browser's content to the agent, so keep it away from logged-in sessions.

No maker's product we track shows it for this yet · in 26 open-source projects.

Playwrightframework · open sourceMicrosoft's MCP server that lets the agent drive a real browser through Playwright's accessibility tree — navigate, click, fill forms, read the page — without screenshots or a vision model.No evidence yet92.2M/wk+17% vs npm

Use it whenThe agent should check the web app it just changed, or reproduce a bug by clicking through it.

Trade-offAccessibility snapshots are large; for coding agents the Playwright team now points to its CLI plus skills as the more token-efficient route.

No maker's product we track shows it for this yet.

Before you choose

How to approach it

Start with nothing extra and note where the agent goes wrong for a week. Add one server or skill for the most common gap, check that it actually changes the result, and remove what you don't use — every MCP server's tool list and every skill's instructions take up context on every task. Prefer servers from the library or tool vendor, and pin versions like any other dependency.

Common mistakes
  • Installing a dozen MCP servers at once, so their tool descriptions crowd the context and the agent picks the wrong tool.
  • Adding a third-party MCP server or skill without reading what it can do, when it runs with your agent's file and shell access.
  • Pointing a docs or wiki service at a private repository without checking where that code is sent and stored.

Other options

Real choices most makers here won't need to weigh.