Basic Agentic Setup — a curated configuration for AI-powered development.
When multiple MCP servers can accomplish a task, prefer in this order:
- Built-in tools (
view,write,edit,bash,grep,glob,ls) — for direct file and shell operations - Domain-specific MCP — use the most specialized server for the task
- General-purpose fallback — Filesystem, Server Memory, etc.
Do NOT call multiple MCP servers for the same task. Pick the best one.
- Use
ctx_batch_executefor research: runs multiple commands, auto-indexes, and enables search. - Use
ctx_searchfor follow-up questions after indexing. - Use
ctx_execute/ctx_execute_filefor data analysis, log processing, and computation in the sandbox. - Do NOT use
ctx_executeto create, modify, or overwrite files. Usewriteoreditinstead.
- When the user asks about a library, framework, SDK, API, CLI tool, or cloud service, use Context7 to fetch current documentation.
- Steps:
resolve-library-id-> pick best match ->query-docswith library ID and question. - Prefer Context7 over web search for library documentation. Training data may be outdated.
- Use Context7 for: API syntax, configuration, version migration, library-specific debugging, setup instructions.
- Do NOT use Context7 for: refactoring, writing scripts from scratch, debugging business logic, code review, general programming concepts.
- Use Serena for intelligent, symbol-level code navigation instead of reading entire files with
view. - Workflow:
get_symbols_overview->find_symbolto locate -> read only the symbol bodies needed. - Use
search_for_patternwhen symbol name or location is uncertain. - Prefer Serena for: symbol navigation, codebase exploration, understanding call graphs, refactoring.
- Do NOT read entire files when Serena's symbolic tools suffice.
- Use Sequential Thinking for complex multi-step reasoning, architectural decisions, and debugging intricate logic chains.
- NOT needed for straightforward tasks (file edits, simple queries, formatting).
- Use Server Memory to persist key facts, decisions, and context across sessions.
- Store: project conventions, team decisions, recurring patterns, important constraints.
- Do NOT store: code that exists in files, git history, ephemeral state.
- Use Filesystem MCP for cross-project file access when needed.
- For files within the current workspace, prefer built-in tools (
view,write,edit) over Filesystem MCP.
- Run Semgrep scans on code changes that involve: auth, user input processing, file system operations, database queries, or crypto.
- Use Semgrep for security-focused code reviews before committing sensitive code.
- Do NOT run Semgrep on every minor change (formatting, comments, config).
- Use Git MCP for repository operations: status, diff, log, branch management.
- For complex git operations (merge, rebase), prefer running git CLI directly via
bash.
- Use GitHub MCP for: issues, pull requests, code search, repository management, and any GitHub API task.
- Use GitHub MCP (not git CLI) when interacting with the GitHub platform (PRs, issues, reviews).
- Use git CLI (not GitHub MCP) for local git operations (commit, branch, stash).
- Use postgres MCP for direct database access: queries, schema inspection, migrations verification.
- Runs with
--access-mode=unrestricted. Be cautious with write operations. - Always use parameterized queries. Never concatenate user input into SQL.
- Before running DDL or destructive queries, confirm with the user.
- Use for analytical queries on columnar data — aggregations, time-series, large-scale analytics.
- Prefer ClickHouse over Postgres for: analytics dashboards, time-series, log aggregation.
- Prefer Postgres over ClickHouse for: transactional data, relational queries, CRUD operations.
- Use Supabase MCP for: database management, auth configuration, storage buckets, edge functions.
- Prefer Supabase over direct Postgres when working within a Supabase project (handles auth, policies).
- For raw SQL on a Supabase database, the postgres MCP may be more direct.
- Use for: live debugging of running web applications — DOM, network, console, performance, Lighthouse audits.
- Use for debugging existing pages. Use Playwright for automating interactions.
- Use for: browser automation, E2E testing flows, automated screenshots.
- Prefer Playwright over Chrome DevTools when: running test sequences, automating repetitive tasks.
- Prefer Chrome DevTools over Playwright when: inspecting a live debugging session, analyzing network traffic, running audits.
- Use for: browsing websites, extracting content without a browser GUI.
- Prefer WebMCP over Playwright for simple content extraction.
- Prefer Playwright over WebMCP for multi-step interactions.
- Postman — full API platform: collections, environments, tests. Use for complete API workflows.
- Postman Code — generate client SDK code from API requests.
- Postman Minimal — lightweight one-off API calls.
- Selection: Minimal for quick tests -> Code for SDK generation -> full Postman for complex workflows.
- Use for GraphQL schema introspection and query execution on a local Apollo server.
- Ensure the Apollo server is running at
http://127.0.0.1:8000before use.
- Use to browse and query OpenAPI/Swagger specifications.
- Useful when: implementing API endpoints from specs, generating types from schemas, validating API contracts.
- Use Figma MCP to access design files: layouts, components, styles, and design tokens.
- Do NOT guess design specifications. Always fetch from Figma when implementing designs.
- Pencil edits
.pendesign files. These files are encrypted — ONLY access them via Pencil MCP tools. - Do NOT use
view,grep, or any other tool to read.penfiles. Always use Pencil MCP (batch_get,batch_design). - Workflow:
get_editor_state->open_document->batch_getto explore ->batch_designto modify. - Max 25 operations per
batch_designcall.
- Use for: research tasks, finding current information, gathering context.
- Prefer Tavily for: research queries, multi-source information gathering.
- Prefer Context7 for: specific library/framework documentation.
- Use for: high-quality web search and content extraction.
- Prefer Exa when: needing high-relevance results, extracting content from specific URLs.
- Prefer Tavily when: needing broader research with configurable depth and images.
| Need | Tool |
|---|---|
| Library/framework docs | Context7 |
| Broad web research | Tavily |
| High-relevance search | Exa |
| Website content extraction | WebMCP or Exa |
| Codebase symbol search | Serena |
| Pattern search in code | Serena or grep |
- Use Docker MCP by default for container management.
- Use Podman MCP only when: Docker daemon is unavailable, rootless containers are required, or user explicitly asks.
- Do NOT run both Docker and Podman MCP for the same task.
- Use for: pod management, deployment status, service inspection, log retrieval, resource scaling.
- Before modifying cluster resources, always show the current state first.
- Use AWS Core MCP for: S3, Lambda, SQS, SNS, DynamoDB, and general AWS API access.
- Use Azure MCP for: resource groups, VMs, storage accounts, and Azure infrastructure.
- Both require appropriate CLI authentication (
aws configure,az login) before use.
- Run Trivy scans on Docker images before deployment, Kubernetes manifests for misconfigurations, and IaC files for security issues.
- Use Trivy alongside Semgrep: Trivy for infrastructure/containers, Semgrep for application code.
- Use to investigate production errors, view stack traces, analyze performance issues, track release health.
- Do NOT use Sentry for local development errors — use Chrome DevTools or terminal output instead.
- Do NOT pass secrets or tokens directly in tool arguments. Use environment variable references.
- When starting a new session, check which MCP servers are relevant to the current task before using them.