| name | nexus-client-onboarding |
| description | First-look workflow for an unfamiliar codebase via the code-nexus MCP server. The right tool order to orient yourself in 30 seconds. Use when you've just landed in a new project and need to know its shape before answering anything specific. |
Codebase Onboarding via code-nexus
When you're dropped into an unfamiliar repo, don't grep blindly. Use the indexed graph in this order.
The 30-second orient
1. reindex(<project-name>) # if not already indexed
2. get_graph_stats() # shape: counts, top-connected modules, languages
3. load_resources(uri: "nexus://project/architecture") # key modules + dependency structure
4. load_resources(uri: "nexus://project/hotspots") # high fan-in/fan-out + dead code count
Now you know:
- What languages are present and how much of each
- The 10 most connected modules (these are the load-bearing structures)
- Which functions get called from everywhere (high fan-in — abstractions)
- Which functions call everything (high fan-out — orchestrators)
Find a starting symbol
If the user asks "how does feature X work?", you don't have a function name yet. Use:
search_code("X", limit: 5)
Pick the highest-scoring result that matches the intent of X (not just keyword overlap). If the top result is a module, drill in:
find_module_hierarchy(entity_name: "TheModule") # behaviours/uses + children
If the top result is a function, get its surroundings:
find_all_callees("the_func") # what does it do?
find_all_callers("the_func") # who depends on it?
get_community_context(file_path: "path/to/file.ex") # structurally coupled files
When the index is stale or missing
reindex is required before anything else when:
- You just cloned the repo
- The container restarted (no auto-create of default collection in v1.2.4+)
- A search returns 0 results that you know should exist
Auto-reindex on queries handles incremental file changes (SHA256-tracked dirty files), but only if the file watcher is set up after a reindex ran at least once.
What nexus://project/overview tells you that you'd otherwise have to compute
load_resources(uri: "nexus://project/overview") gives you:
- File count, total chunks, graph node count
- Language breakdown (chunks per language)
- Entity types (function/module/class counts)
This is not in get_graph_stats — overview is more readable, get_graph_stats is more granular (top-connected lists, edge counts).
Multi-project gotcha
Each project lives in its own Qdrant collection (nexus_<name>). When you reindex(other-project), the active collection switches. Subsequent queries hit other-project only. To switch back, reindex(<original>) (data is preserved per collection — just switching which one is active).
If you need both at once, that's not supported — restart the workflow per project.
