Consumes much more tokens than letting agent explore by itself

[freakynit]

Version

0.9.0

Platform

macOS (Apple Silicon)

Install channel

GitHub release archive / install.sh / install.ps1

Binary variant

standard

What happened, and what did you expect?

This actually consumes more tokens than letting agent explore itself. Have tested it twice earlier (around 2 months back), and tested again just now.

Coding agent harness: pi
Test repo: https://github.com/chattocorp/chatto
Query:
-- 1. Use codebase-memory-mcp and tell me how authentication works in current project
-- 2. Without using codebase-memory-mcp tell me how authentication works in current project

Results

1. With codebase mcp: 74K input tokens
2. Without: 43K

Reproduction

1. Clone the repo
2. Integrate MCP
3. CD into the cloned repo directory
4. Start PI agent
5. Ask question

Logs

Diagnostics trajectory (memory / performance / leak issues)

Project scale (if relevant)

No response

Confirmations

• I searched existing issues and this is not a duplicate.
• My reproduction uses shareable code (a dummy snippet or a public OSS repository), not proprietary code.

[github-actions]

Thanks for the report! To move this forward we need a bit more so we can reproduce it ourselves:

• the codebase-memory-mcp --version you're on
• the exact steps or command you ran
• a public repo (or a small dummy snippet) that shows the problem — please don't paste proprietary code

Once that's here we'll pick it straight back up. Heads-up: issues left awaiting-reporter are automatically closed after a few weeks of silence, but a comment reopens the door anytime.

[DeusData]

Thanks for publishing the repository, prompt, client, and token totals. The comparison is useful, but aggregate input tokens alone do not yet isolate a server defect: the result depends on the client's tool descriptions, system prompt, model, conversation state, indexing state, tool-call count, pagination, and whether the two runs produced equivalent answers.

Please add fresh-session transcripts for both runs with the exact pi and model versions, tool calls and responses, final answers, and token accounting split by system/context/tool input. Also confirm whether the project was already indexed before the MCP run. That will let us distinguish excessive MCP payloads from the client repeatedly invoking tools or comparing unequal answers. Marking awaiting-reporter for that controlled A/B evidence.

[freakynit]

@DeusData Everything mentioned is open/accessible. You can do it yourself as well. I just wanted to highlight this by giving enough info for you folks to verify yourself.

I'm personally not using it any more.

Thanks..