Skip to main content

1.8 Code Intelligence for AI-Native Workflows

These tools help your agent find answers without reading entire files. Choose the type that fits your question, install a tool, and add the instructions below to AGENTS.md.

Understand​

These four categories of tools answer different types of questions:

FamilyQuestion it answers
Indexing"Where is the code that handles authentication?"
LSP"Where is this symbol defined? What calls it? What is its type?"
Data flow"Does untrusted input reach this sink? What breaks if this changes?"
Docs RAG"How do I use this dependency at the version installed in my project?"

Each category includes a Solo Dev option for individual developers and an Org option for teams that need shared infrastructure. Team options may require hosting or a paid license. This quick start uses the Solo Dev options.

Problems it solves​

These tools can reduce the tokens your agent spends finding relevant code:

QuestionBaseline (Grep + file reads)With the right family
"Where is the code that handles authentication?"~12 tool calls, ~24k tokens on discovery1 Indexing call, ~400 tokens
"What functions call verifyJWT?"~8 tool calls, ~15k tokens1 LSP call, ~300 tokens
"Does HTTP input reach os.system in this codebase?"Not answerable by Grep at all1 Data flow call, taint trace with file/line hops

Building an index or configuring retrieval takes some setup, but can make later searches cheaper. For measurements from a real repository, see the paired lab runs in the labs/ directory of the comparison repository. Each scored tool has a fixed setup and prompts for comparing it with the baseline.

Base implementation (Solo Dev)​

Choose one category that matches the questions you ask most often. Install its Solo Dev tool and add the provided instructions to AGENTS.md (or CLAUDE.md). Then ask a question and check which tools the agent calls.

FamilyInstallInstructions to add to AGENTS.md
Indexingpipx install 'cocoindex-code[full]' ccc init claude mcp add cocoindex -s user -- ccc mcp"When looking for code by what it does rather than by name, use the cocoindex MCP search tool."
LSPclaude mcp add serena -s project -- uvx --from git+https://github.com/oraios/serena serena start-mcp-server --project $(pwd)"For symbol / reference questions use serena find_symbol, find_referencing_symbols."
Data flowpip install semgrep claude mcp add semgrep -s user -- semgrep mcp"For taint / reachability questions use semgrep MCP tools; pattern authoring looks like the code."
Docs RAGclaude mcp add context7 -s user -- npx -y @upstash/context7-mcp --api-key YOUR_API_KEY"For third-party library docs use context7 query-docs, pinned to the version in the lockfile. Verify the Source: line matches the ref you asked for."

Then ask a question the tool is designed to answer. Check which tools the agent calls: did it use the MCP tool or fall back to Grep? If it fell back, make the instructions more explicit (**always** call ...).

Next steps​

  • Setup guides: Indexing · LSP · Data flow · Docs RAG. Each page covers installation for individual developers and setup options for teams.
  • Compare tools within a family: open the comparison site and go to research/<tool>/ for the scored capability matrix and per-cell reasoning. labs/<tool>/ holds pinned setups and paired baseline-vs-tool evaluations.