Codegraph consumes too much Claude context

[starrylistener]

In many cases I only want to locate relevant files/symbols or get a high-level overview, but the tool returns large source snippets from multiple files. This
quickly fills the conversation context and makes it harder to continue working in the same session.
Claude’s context window.
Is there a recommended way to reduce context usage?

[colbymchenry]

Thanks for raising this. Quick question first: what CodeGraph version are you on (codegraph --version), and which tool is producing the large output — codegraph_explore?

Asking because a lot of output-sizing work landed recently:

• 0.9.8 turned on adaptive sizing by default — codegraph_explore now keeps only the methods on your flow path in full and collapses redundant/off-path code to signatures.
• 0.9.9 made it size the response to the answer rather than the file count, and stop emitting partial methods.
• The next release adds Read-style offset/limit to codegraph_node so a file can be read in slices.

If you're below 0.9.9, upgrading alone should cut the volume noticeably.

For your two specific cases, there are lighter-weight tools than codegraph_explore (which intentionally returns full source so the agent doesn't have to re-read files):

• Just locating a symbol → codegraph_search returns its kind, location, and signature with no body.
• A high-level structural overview of a file → codegraph_node with symbolsOnly: true returns the file's symbols without their source.

And if your agent keeps reaching for codegraph_explore when you'd rather it didn't, CODEGRAPH_MCP_TOOLS=search,node,callers,callees,impact trims the exposed tool set so the heavy one isn't offered.

Let me know your version and roughly what you were asking the agent to do, and I can tell you whether what you hit is already addressed or something we should tune further.

[starrylistener]

Thanks for the detailed explanation.

I’m on CodeGraph 0.9.9, and yes, the large output mainly comes from codegraph_explore.

My use case is usually asking Claude Code broad questions like:

• “How does this indexing flow work?”
• “Where is this feature implemented?”
• “Explain the structure around this module.”

So I understand that codegraph_explore is intentionally returning source so the agent does not need to re-read files. The issue is that for broad
exploration questions, it can still take a noticeable amount of context, even on 0.9.9.

Based on your suggestions, I’ll try using the lighter tools more explicitly:

• codegraph_search when I only need symbol location
• codegraph_node with symbolsOnly: true for file/module overview
• possibly limiting exposed tools with CODEGRAPH_MCP_TOOLS=search,node,callers,callees,impact

One thing that might still help is a lighter exploration mode for codegraph_explore, or a way to ask it to return “navigation first, source later.”
For example, first return the relevant files/symbols/call path, then only expand source for selected methods.

That would make it easier to use CodeGraph for early-stage exploration without immediately spending a lot of Claude context.