Агент Pi. Полный гайд по агенту. LSP, MCP, extentions, skills. Настраивам агент и улучшаем код.

Summary

The video is a practical guide to configuring Pi, a minimalist coding agent that runs in the terminal. The presenter prefers Pi’s small core over feature-heavy, all-in-one agents: Pi provides the basics, while language support and other capabilities can be added through extensions, MCP servers, skills, and packages.

Installation and configuration

• Pi can be installed with a shell command on macOS, Linux, or Windows.
• The main model configuration file is models.json. It stores provider details such as the model name, base URL, and API key, and can be configured for cloud providers or local model servers.
• The presenter also covers Pi’s account, settings, MCP, and instruction files. He recommends using AGENTS.md for general agent guidance and adding project-specific instructions within each project, rather than changing Pi’s standard system prompt without a clear reason.

Understanding and working with code

• LSP (Language Server Protocol): Connects language servers to Pi through an LSP extension, allowing the agent to query code structure—including symbol references, definitions, types, and errors—instead of relying only on text search. The presenter demonstrates a reference lookup and recommends installing servers for the languages used in a project.
• Codemode: Lets Pi write and run short JavaScript scripts in a sandbox to call tools, including MCP tools, asynchronously. The presenter says this can speed up tasks that would otherwise require multiple separate tool calls.

MCP tools

The presenter demonstrates or recommends several MCP integrations, noting that their usefulness depends on the project:

• Context7: Provides current library and framework documentation, helping reduce outdated API assumptions.
• SearXNG: A self-hosted metasearch service that searches multiple sources and returns results through an API. The presenter distinguishes it from Context7: SearXNG is for general web searches, while Context7 is for library documentation.
• Chrome DevTools MCP: Lets the agent inspect and interact with a running website, including page elements, console errors, network requests, screenshots, and browser storage. The presenter recommends it for web development.
• Playwright MCP: Focuses more on browser automation and testing user-facing behavior.
• Postgres MCP: Can help inspect queries, explain plans, indexes, migrations, and database performance.
• RAG MCP: The presenter runs a personal retrieval service backed by Qdrant, a vector database. It has endpoints for adding, searching, and deleting stored information. He describes RAG as useful for finding semantically relevant material in a large knowledge base, rather than only exact text matches.
• Postman MCP: The presenter found it offered little benefit for his API work because he felt the agent could already handle much of that work without the integration.

Pi extensions and customization

The guide covers several extensions:

• LSP extension: Connects Pi to language servers.
• PDR: Adds another working directory to the agent’s accessible project area.
• FFF: Adds fast file-search tools that Pi can use alongside commands such as find or grep.
• Neovim integration: Lets the user send the current file or code location to Pi and ask questions about it in context.
• Sub-agents: Support delegating work across models or using different models for different tasks. The presenter says he does not currently use this extension.
• Apply Patch: Lets the user review diffs before applying changes.
• Toolwatch: Adds permission controls for restricting potentially dangerous actions. The presenter notes that Pi is permissive by default.

Sessions, context, and project guidance

• Pi saves sessions as files and can resume previous work.
• The presenter explains commands for starting a new session, resuming or branching through the session tree, forking or cloning a session, compacting context, naming a session, and exporting it.
• Compaction summarizes older conversation content to make room in the model’s context. According to the presenter, the original conversation remains available in the session history.
• AGENTS.md files provide persistent project-specific context and rules, such as framework and language versions, coding conventions, or constraints. Instructions in nested project folders can supplement broader project guidance.
• Skills are reusable Markdown instructions for recurring workflows, such as migrations, deployment, or testing. The presenter recommends creating skills for tasks specific to one’s projects instead of installing large collections indiscriminately.
• Packages bundle Pi customizations—such as extensions, skills, prompts, and themes—for sharing with colleagues or moving between computers. The presenter mentions publishing his own package on GitHub.

Overall assessment

The video’s central recommendation is to start with Pi’s minimal core and add only the tools a workflow needs. LSP and up-to-date documentation are presented as especially helpful for coding. Browser and database MCPs can be useful for relevant projects, while some integrations may add little value. The presenter also demonstrates parts of his own setup, including terminal-based project work, LSP lookups, MCP connections, and session management.

Main speaker and source

The video features one presenter, who explains and demonstrates his personal Pi configuration and shares his preferences and recommendations.