Abhay Ganti

CS + Math @ NUS| Singapore | Interested in AI/ML, LLMs, agents and full-stack dev

[Abhay Ganti @silvermango9927]

@amirihayes_ woah, very cool! how'd u get started with research?

[Abhay Ganti @silvermango9927]

Updated audit-mcp from maintainer feedback.

Big lesson: MCPs can get worse as you add tools/descriptions. Overlap grows, flows blur, and agents stop picking tools.

Now reports focus on drift, grouping, rewrites, fix plans, and PR checks.

Looking for MCPs to audit. DMs are open

[Abhay Ganti @silvermango9927]

@harleyfoote_ Hey, noticed this MCP server. I created mcplens-cli, to audit MCP servers and return reports for execution. Mainly looking for vague tool descriptions, overlap, and default-surface bloat and optimizing it by reducing token usage. Would love for you to use it if you're building!

[Abhay Ganti @silvermango9927]

@AkankshaNarula8 hahaha sure! waiting for it

[Abhay Ganti @silvermango9927]

I'd really love for MCP authors to use it and give me any feedback
NPM : https://t.co/avmuFn2AD8
Github : https://t.co/pAScIqIQID

#mcp #agents

[Abhay Ganti @silvermango9927]

Launching mcplens-cli.

MCP servers often fail because agents can’t tell which tool to use, the surface is noisy, or responses are bloated.

mcplens-cli audits MCP tool activation locally and helps generate lean MCP servers from OpenAPI specs.

Link below.

[Abhay Ganti @silvermango9927]

@belindazli @AnthropicAI hey! not sure if you'll take a look at this, but im a y1 cs student, and i'm looking to learn interpretability, and i started doign the ARENA 3.0 course? how would you recommened i learn? thank you so much!

[Abhay Ganti @silvermango9927]

@punitweb @KanakDaswani yeah, my bad for not clarifying. It's generally meant for MCPs that you make, because tool count and response bloat is generally a big issue for efficinecy, and maximising value out of agent context windows. so i was wondering if you've encountered issues of that sort.

[Abhay Ganti @silvermango9927]

@crack3nnn yeah, i'm building a CLI that essentially helps trim down response bloat by 76-85% by my initial runs on stuff like Slack/Notion/Jira/Github etc. but i'm trying to understand pain points, because too many endpoints/tools harms agents more than benefits

[Abhay Ganti @silvermango9927]

@KanakDaswani @punitweb just to see if optimizing it ould solve pain points

[Abhay Ganti @silvermango9927]

@KanakDaswani @punitweb Curious: when you generated this MCP server, did the raw tool surface work as-is?

I’m testing whether the real pain is not generation, but post-generation cleanup:
tool pruning, response trimming, unsafe endpoint hiding, better task-level tools etc. cheers!

[Abhay Ganti @silvermango9927]

@ericzakariasson just to understand if optimizing it would solve some pain points. cheers!

[Abhay Ganti @silvermango9927]

@ericzakariasson Curious: when you generate mcp servers, do the raw tools surface work as-is?

I’m testing whether the real pain is not generation, but post-generation cleanup:
tool pruning, response trimming, unsafe endpoint hiding, better task-level tools, etc.

[Abhay Ganti @silvermango9927]

@AmoebaFreaks @browserOS_ai Curious: when you generate MCP servers, do the raw tool surface work as-is?

I’m testing whether the real pain is not generation, but post-generation cleanup:
tool pruning, response trimming, unsafe endpoint hiding, better task-level tools, etc.

[Abhay Ganti @silvermango9927]

#mcp #agents #ai #tools #mcpservers #agentic

[Abhay Ganti @silvermango9927]

i'm quite inactive on twitter, but i've been building something.
OpenAPI → MCP is easy now.

Raw generated MCP servers can still be bad for agents:
- too many tools
- unsafe/admin endpoints exposed
- huge response payloads
- noisy schemas
- agents picking the wrong tool

[Abhay Ganti @silvermango9927]

I’m looking for 5 people who will let me run this on a real OpenAPI spec or MCP server and send back a before/after diff.

Not selling anything. I’m trying to learn if this is a real pain or just a neat tool.