AI tools
Tools for the jobs around a coding agent
Seven tools for the work around a coding agent: finding skills and MCP servers, preparing a repository, giving an agent a browser or a design library, and checking what it actually did.
A coding agent arrives with chores attached. Someone has to find the skills and MCP servers it will use, make the repository readable, decide what it may touch and check its work afterward. Each tool here takes one of those chores. Most of them also ask you to trust something: a stranger's skill, a signed-in browser, a paid library. Every entry says what to check before you hand anything over.
Find a component
Two catalogs of skills, plugins and MCP servers. Compare how clearly each one shows where a listing came from.
agentskills.codes
Why it belongs here
agentskills.codes is a searchable registry of agent skills, browsable by client, category and popularity, with a copyable install command on every entry for Codex, Claude Code, GitHub Copilot and others. The site says it rescans listings daily. That copy button is both the convenience and the risk. A high popularity count doesn't tell you a skill is safe or maintained, so read the source and its permissions before you paste anything.
Read full NoteNote published Aug 11, 2026AgentMatter
Why it belongs here
AgentMatter catalogs public skills, plugins, MCP servers and prompt collections, and it puts the dull but important details right beside each listing: the GitHub repository, the owner, the license and the install steps. When the thing you're adding may run with your agent's permissions, seeing where it came from before you copy it is the right habit. Its entries, licenses and compatibility claims weren't verified for its Note.
Read full NoteNote published Aug 26, 2026
Prepare the repository
Before an agent touches the code, check that the instructions it depends on actually exist.
Agent-Ready Repo Auditor
Why it belongs here
"Agent-ready" usually gets discussed in vague terms. Agent-Ready Repo Auditor turns it into a checklist. Run it as a CLI or GitHub Action and it looks for written instructions, issue intake and verification commands, then writes a report. It reads public GitHub data without cloning or running your code, and a private or truncated repository fails outright instead of getting a partial score. The score grades documentation. It says nothing about code quality or security, or about leaving an agent to run alone.
Read full NoteNote published Jul 21, 2026
Give an agent a connection
One hands an agent a real browser, the other a paid design library. Settle access and terms before either.
Causeway
Why it belongs here
Causeway is a single Rust binary that sits between an MCP client and a real Chromium browser, turning tool calls into DevTools Protocol commands for navigation, clicks, forms, screenshots and network or console inspection. The design is refreshingly simple, and that simplicity cuts both ways: an agent connected to a signed-in browser can reach every session open in it. Read its source, configuration and permissions before you point it at real accounts.
Read full NoteNote published Aug 11, 2026ScreensDesign MCP
Why it belongs here
A paywall screenshot doesn't show how an app talked someone into reaching it. ScreensDesign MCP lets your assistant search recorded mobile app flows in order, so you can ask for the screens before a subscription offer and follow links to the recordings. It comes with ScreensDesign Pro, listed at $39 a month or $299 a year. Its revenue estimates don't prove a paywall works, recordings age as apps change, and the terms forbid building datasets or services from what you pull out.
Read full NoteNote published Oct 15, 2026
Show and check the output
Render a response while it streams in, then trace the steps the agent took to produce it.
Markstream
Why it belongs here
Most Markdown renderers are judged after the last token lands. Markstream is built for the messy seconds before that, when a code fence is still open and a table is half written. It renders streaming AI responses, SSE and WebSocket output as they arrive, including code blocks, Mermaid diagrams and KaTeX, with guides for Vue, React, Svelte, Angular, Nuxt and Next.js. Its team says that if you only need a finished document turned into HTML, marked or markdown-it is simpler. Its Note covers the documentation, not an install or performance test.
Read full NoteNote published Jul 21, 2026AgentInspect
Why it belongs here
An agent can hand back a plausible answer after skipping a step it was meant to take. AgentInspect records a TypeScript agent's nested steps, tool calls, model metadata and errors in a local trace, and its CI checks can require that particular tools actually ran. The README is careful to say a passing check isn't proof the task succeeded, and that redaction for shared bundles is best effort. The package wasn't installed or run for its Note; its synthetic example is the place to start.
Read full NoteNote published Sep 19, 2026
How these projects were chosen
Every pick has a published Project Note and covers a different step, so the groups follow the workflow instead of sorting tools from best to worst. Most of these Notes draw on public pages, READMEs and documentation without installing the software, and each entry says what that leaves open. Popularity scores, sponsorship and clicks play no part in what's included or in the order.
For developers who already work with an agent such as Codex, Claude Code or Copilot and want one specific tool for one step of that work, along with the permission questions to settle before installing it.
A review can leave the picks as they are. When a published selection or its guidance changes, a short note here explains why.