source browsing mcp

Read less, resolve faster, keep context tight.

Use `eye` when an agent needs exact code navigation, not more file churn.

The docs site focuses on setup, client wiring, bounded navigation patterns, and the five MCP tools that keep large repositories inspectable.

Move from install to first useful query in one pass.

01

Install the local runtime

Prepare Node.js, Corepack, Universal Ctags, and ripgrep before wiring the MCP server.

02

Build the stdio entrypoint

Run pnpm install and pnpm run build once, then point your client at the dist entry.

03

Navigate with bounded reads

Start with structure, read exact source ranges, then reuse symbolId for follow-up lookups.

A compact MCP surface with clear job boundaries.

get_project_structure

Map a repository without walking every folder by hand.

read_source_range

Read exact line-centered context with numbered output.

query_symbol

Resolve definitions, references, and bounded context through one surface.

refresh_index

Refresh the local cache for the whole root or a narrow scope.

get_index_status

Inspect cache generation, counts, and readiness without writes.

Keep the entrypoint boring so the docs can stay precise.

claude mcp add --scope project eye -- node /absolute/path/to/eye/dist/index.js

Current product limits are part of the public contract.

One project root at a time, resolved automatically or with projectRoot.
Lazy local indexing instead of a background daemon.
Index-first symbol lookup with explicit fallback search when needed.

Need the full path?

Walk from install through prompt patterns, then use the tool reference and FAQ as the stable public contract.

Read the FAQ