@maven_hq: MCP has a problem. This is just one of the many problems with MCP. If you want to learn more about it from industry experts, check out the link in my bio. #maven #ai #llms #mcp
this is why tool catalogue size matters so much. clients re-send every tool definition on every turn. Owl Browser runs a 20-tool profile by default instead of 80 to keep that context tax down.
2026-08-18 13:14:09
0
flapfinned :
Need a mcp for my mcp
2026-01-23 15:56:42
20
olayinka_ai :
Claude tool discovery
2026-03-22 06:54:57
1
fr3klz :
CLI ftw
2026-03-21 07:05:29
0
rickdeedoo :
I like when you got to the solution
2026-03-05 21:23:40
0
Luria :
Just... use chroma
2026-03-06 13:34:38
0
Gambit :
but you have an orchestration agent that uses other agents with their own context to use different tools that send their response to the orchestration agent which uses a summarised context
2026-01-24 06:55:04
0
Trystan - Product engineer :
RAG to retrieve MCP ? no ?
2026-01-23 21:38:18
0
neverexceptalways :
I keep them all off until I want to use one
2026-01-24 19:25:09
0
Fernando :
lazy loading for mcp is already available for some implementations
2026-03-13 20:12:54
0
Mack Solomon :
seems like a easy problem to solve
2026-02-15 18:59:24
0
Mk11233 :
Claude code fixed this in their updates last week. They now delay load if they take too much context which is great
2026-01-24 20:36:01
0
plxnet :
Just limit the tools the agent has access too...
2026-01-23 17:42:23
0
Hasan :
Video just cuts off
2026-01-23 19:39:32
4
JeyGDiff :
Seems like a great opportunity for the harness to enlist a subagent with a small problem statement, gather MCP tools and return a succinct list of useful ones given the problem statement back to the main context agent.
2026-01-24 04:00:57
0
Bloody Kheeng :
it gets worse if a tool has a bug it crashes lol 😂😂
2026-02-10 06:46:21
0
toby brooks :
And it all be fixed in the future
2026-10-03 10:24:48
0
QuantumCoherence :
2026-06-08 17:04:19
0
` :
Just use a subagent with only skills loaded and use the main agent for the tasks and orchestration, I'm saying this with only an idea in mind as for how to fix this issue
2026-01-26 09:05:03
0
Aaron Wright :
Context rot is real so I moved my setup to gen.pro where I can test new models in one place without rebuilding my entire tool schema every time.
2026-09-29 11:43:37
0
Kuro :
Just enable and disable them when you need them. Though this removes autonomy.
2026-01-24 11:22:50
0
zach :
This was solved already
2026-01-24 06:12:52
0
alienwatchers 🟦 :
u mean vector to get from rag then direct the llm?
2026-01-24 02:12:31
0
To see more videos from user @maven_hq, please go to the Tikwm
homepage.