Инструменты

GitHub's official MCP server: lets agents read and write repositories and code, manage issues and pull requests, and check Actions workflows and code-scanning alerts. Use the hosted remote server or run it locally in Docker; needs a personal access token or OAuth.

Anthropic's role-based skill collection covering sales, finance, legal, HR, support, marketing, product, design, operations and data analysis: over a hundred skills you can install one at a time. Works best with CRM, data warehouse and other connectors.

Anthropic's knowledge-work plugin marketplace, bundling skills, commands and connectors by role: sales, support, product, marketing, legal, finance, HR, data and more, plus many partner plugins. Built mainly for Cowork; also works in Claude Code.

Google's open-source database MCP server with ready-made query and SQL tools for dozens of sources, including PostgreSQL, MySQL, BigQuery, Spanner and MongoDB. Define controlled custom queries in tools.yaml.

Skills from security firm Trail of Bits: 40+ plugins for code auditing, CodeQL and Semgrep static analysis, differential review, smart-contract and cryptography audits, fuzzing and more. Also installable as a Claude Code or Codex plugin marketplace.

Cloudflare's official remote MCP servers: manage Workers and bindings such as KV, R2 and D1, and query logs and observability, builds, DNS analytics and Radar data. Authorizes with Cloudflare OAuth; some features need a paid plan.

A four-phase root-cause debugging method: investigate the root cause, compare against working examples, test one hypothesis at a time, then fix the root cause. Guessing at code changes before locating the cause is not allowed. From the Superpowers workflow.

Before writing code, asks questions one at a time to clarify intent, constraints and success criteria, turning a vague idea into a reviewable design; implementation starts only after you approve it. A core Superpowers skill.

Once requirements or a spec exist and before touching code, writes multi-step work as a fine-grained implementation plan listing the files, interface signatures and tests to verify each step. A core Superpowers skill.

Forces red-green-refactor: write a test and watch it fail, then write the minimum code to pass. Any code written without a verified failing test gets deleted and redone. From the Superpowers workflow.

Sets up an isolated workspace before feature work: uses the agent's built-in worktree tool when available, otherwise falls back to git worktree, then installs dependencies and runs baseline tests to confirm a clean start. From Superpowers.

Follows an implementation plan by dispatching a fresh subagent for each task, reviewing each for spec compliance and code quality, then doing a final overall review. From Superpowers; use with its code-review and branch-finishing skills.

Requires the agent to run the full verification commands and read the output before claiming work is done, fixed or passing, backing claims with evidence and ruling out unverified success reports. From the Superpowers workflow.

Before finishing a task or merging, dispatches an independent reviewer subagent over a git commit range to check the work against the plan and requirements, then handles feedback by severity: critical, important, minor. Part of obra/superpowers.

A disciplined process for hard bugs and performance regressions: build a reliable reproduction loop, reproduce and minimize, form hypotheses, instrument to verify, then fix and add a regression test, redacting secrets in output. By Matt Pocock.

Before you build, the agent grills you on your plan or design, offering a recommended answer each round, until every decision branch is resolved. It calls the grilling skill from the same repo, so install both.

Test-driven development in vertical slices: agree on test seams with the user first, verify behavior only through public interfaces, and avoid listed anti-patterns such as implementation coupling and tautological tests. From Matt Pocock's skills collection.

A complete harness for coding agents: dozens of subagents, nearly 300 skills, hooks, rules, memory and continuous learning, plus the AgentShield security scanner. Built mainly for Claude Code, with Codex support.

Self-improving agent from Nous Research with built-in memory and a learning loop that turns experience into skills. Works with any model provider; the hermes command has a terminal UI, and a gateway connects it to Telegram, Slack and more and runs scheduled tasks.

Behavior guidelines drawn from Karpathy's observations of how LLMs go wrong when coding: think before acting, keep it simple, make surgical changes only, and drive work by verifiable goals. Can also be used directly as a CLAUDE.md.

Four guidelines based on Karpathy's observations of common LLM coding mistakes: think before coding, prefer simplicity, make only necessary changes, and work toward verifiable goals. Ships as a plugin skill or works as a CLAUDE.md. Community-made, not by Karpathy.

Open-source terminal coding agent that switches freely between model providers. Includes a build agent that edits code, a read-only plan agent for analysis, subagents, and a desktop app. Formerly sst/opencode.

Guides users through co-writing proposals, technical specs and decision docs in three stages: gather context, refine section by section, then test with a fresh session that lacks the context to catch what readers won't understand. Official Anthropic skill.

After giving advice, a draft or a data analysis, adds 2-3 targeted follow-up questions that prompt the user to check key facts, test the reasoning and fill in missing context. Fires at most once per conversation. Official Anthropic skill.