What problem does this solve?
When Claude Code explores a codebase, it spawns a predefined Explore Subagent that only has access to a fixed set of tools: Grep, Glob, and Read.
This means that even when codebase-memory is installed and configured as an MCP server, the Explore Subagent is completely unaware of it — and exploration happens without leveraging codebase-memory's capabilities. The result is slower, less context-aware exploration that ignores a tool specifically designed for this purpose.
Concrete example:
Claude Code is asked to implement a new feature. It spawns the Explore Subagent to understand the codebase structure. Because the subagent only has Grep/Glob/Read, it manually scans files instead of querying codebase-memory — defeating the purpose of having codebase-memory installed.
Proposed solution
When codebase-memory is integrated into a host system (e.g. Claude Code), it should automatically generate and register a custom agent definition file — e.g. a explore-agent.md or equivalent — that:
- Defines a drop-in replacement (or extension) of the default Explore Subagent
- Includes codebase-memory's MCP tools in the agent's available tool set
- Follows the host system's agent definition format (e.g. Claude Code's
.claude/agents/ directory)
This way, when Claude Code decides to spawn an Explore Subagent, it picks up the codebase-memory-aware version instead of the built-in default.
Why this matters
| Without this feature |
With this feature |
| Explore Subagent ignores MCP tools |
Explore Subagent leverages codebase-memory |
| Redundant file scanning |
Semantic/indexed codebase queries |
| MCP integration is "opt-in per prompt" |
MCP integration is seamless & automatic |
Additional context
- Claude Code stores custom subagent definitions under
.claude/agents/
- The generated agent file could be created during
codebase-memory init or on first MCP connection
- Should be idempotent — re-running init updates the file rather than duplicating it
Alternatives considered
No response
Confirmations
What problem does this solve?
When Claude Code explores a codebase, it spawns a predefined Explore Subagent that only has access to a fixed set of tools:
Grep,Glob, andRead.This means that even when codebase-memory is installed and configured as an MCP server, the Explore Subagent is completely unaware of it — and exploration happens without leveraging codebase-memory's capabilities. The result is slower, less context-aware exploration that ignores a tool specifically designed for this purpose.
Concrete example:
Proposed solution
When codebase-memory is integrated into a host system (e.g. Claude Code), it should automatically generate and register a custom agent definition file — e.g. a
explore-agent.mdor equivalent — that:.claude/agents/directory)This way, when Claude Code decides to spawn an Explore Subagent, it picks up the codebase-memory-aware version instead of the built-in default.
Why this matters
Additional context
.claude/agents/codebase-memory initor on first MCP connectionAlternatives considered
No response
Confirmations