Using Gkg

using-gkg — Knowledge Graph Exploration

gkg is a knowledge graph tool for codebases that indexes symbols, definitions, and relationships, then enables fast search and contextual lookup. Use it when you need to understand a codebase structure, find definitions, trace usages, or explore cross-module relationships.

What gkg Is

gkg builds a semantic index of a codebase — symbols (functions, types, variables, imports), their definitions, usages, and relationships. Unlike grep or rg, which search text, gkg understands code structure.

When to use gkg:

• Finding all usages of a symbol across a large codebase
• Tracing the definition chain (where is MyType defined? what implements it?)
• Understanding cross-module relationships
• Exploring unfamiliar codebases quickly

When to use grep/rg instead:

• One-off text search in a few files
• Pattern matching on specific keywords or strings
• No preprocessing time available (gkg indexing takes seconds)

Step 1 — Index a Repo

Index a codebase to build the knowledge graph:

gkg index /path/to/repo

This scans the repo, builds a symbol index, and writes it to .gkg/ (in the repo root or a temporary location). Indexing is fast (typically <10 seconds for moderate repos).

Example:

gkg index /Users/joe/dev/minibox

After indexing, you can search the same repo without re-indexing.

Step 2 — Search Symbols

Search for a symbol, function, type, or pattern:

gkg search "<query>"

Returns all matches with file paths and line numbers.

Examples:

# Find all definitions and usages of the function `deploy`
gkg search "deploy"

# Search for a type definition
gkg search "type Agent struct"

# Find imports of a module
gkg search "import mbx"

Step 3 — Get Context on a Symbol

Look up the definition and usage context for a specific symbol:

gkg context <symbol>

Returns the symbol definition, declaration location, and nearby code context.

Example:

gkg context Agent

MCP Access via mcpipe

gkg is accessible via MCP (through the mcpipe local proxy). The SSE endpoint is:

http://localhost:27495/mcp/sse

This allows Claude Code to query the knowledge graph programmatically without running shell commands. The endpoint supports:

• gkg.index — trigger indexing
• gkg.search — search for symbols
• gkg.context — get definition context

To use via MCP, you do not need to run gkg manually; the plugin will invoke it through the SSE endpoint.

Workflow Example

Scenario: Exploring the minibox codebase

1. Index the repo:

   gkg index /Users/joe/dev/minibox
   

2. Search for the main Agent type:

   gkg search "struct Agent"
   

   Output: file paths and line numbers of all Agent definitions and usages.

3. Get context on the Agent type:

   gkg context Agent
   

   Output: the struct definition, its fields, and usage patterns.

4. Search for callers of a specific function:

   gkg search "fn provision"
   

   Output: all references to the provision function.

Performance & Limits

• Indexing: typically <10 seconds for repos up to ~100K lines
• Caching: index is cached locally; re-run gkg index to refresh
• Search: returns all matches; filter output if needed

Troubleshooting

gkg: command not found

Ensure gkg is installed:

which gkg

If not found, install via mise:

mise install gkg

Index is stale

Re-index the repo:

rm -rf .gkg/
gkg index /path/to/repo

MCP endpoint not responding

Verify mcpipe is running:

curl http://localhost:27495/mcp/sse -v

If not, start mcpipe as documented in the plugin configuration.

Additional Resources

• gkg docs: Run gkg --help for CLI reference
• mcpipe: Local MCP proxy configuration and setup