Problem
When an agent starts a new session on an unfamiliar codebase, it currently needs to:
scan the workspace
list all files
detect frameworks
- Read several entry point files manually
This costs 10-20k tokens just for orientation.
Proposed Change
Add a get_architecture MCP tool that returns a structured summary in one call:
{
"languages": {"python": 342, "typescript": 89},
"frameworks": ["fastapi", "react"],
"entry_points": ["src/main.py", "src/server.ts"],
"packages": ["src/api", "src/models", "src/services"],
"routes": [{"method": "GET", "path": "/users", "handler": "get_users"}],
"hotspots": ["src/models/user.py (47 dependents)"],
"total_symbols": 1842,
"adrs": [] // if ADR feature implemented
}
Why This Matters
An agent that calls get_architecture first has enough context to ask precise questions instead of exploring blindly. Target: orientation in 1 tool call, <1k tokens output.
Problem
When an agent starts a new session on an unfamiliar codebase, it currently needs to:
scanthe workspacelistall filesdetectframeworksThis costs 10-20k tokens just for orientation.
Proposed Change
Add a
get_architectureMCP tool that returns a structured summary in one call:{ "languages": {"python": 342, "typescript": 89}, "frameworks": ["fastapi", "react"], "entry_points": ["src/main.py", "src/server.ts"], "packages": ["src/api", "src/models", "src/services"], "routes": [{"method": "GET", "path": "/users", "handler": "get_users"}], "hotspots": ["src/models/user.py (47 dependents)"], "total_symbols": 1842, "adrs": [] // if ADR feature implemented }Why This Matters
An agent that calls
get_architecturefirst has enough context to ask precise questions instead of exploring blindly. Target: orientation in 1 tool call, <1k tokens output.