Best Local MCP servers for coding support?

Posted by Dundell@reddit | LocalLLaMA | View on Reddit | 4 comments

I'm coming from a place where I understand some of them for web searches and adding features to your LLM. There was recently an Aider Polyglot benchmark for GPT-OSS 120B using Aider's Repomap to a 78.7% which is very desirable as a local model. This Repomap can be a local MCP server and just reading into it as a RooCode user, how to set it up and what is required. I have no idea further what is a normal local MCP peopl .use for coding support that's easy to setup for users?

Reply to Post

4 Comments

[-]

Dundell@reddit (OP)

I found [pdavis68/RepoMapper](https://github.com/pdavis68/RepoMapper) and have spent probably 10 hours today just reworking and trial/error for my local setup. Soi far it looks accurate for my 4 personal github projects tested and does seem to help, but I will need to test my fork out more on my end: [https://github.com/ETomberg391/RepoMapper](https://github.com/ETomberg391/RepoMapper) So there was that and a lot of talk about Context7 which seemed interesting I'll have to implement tomorrow, although some people claimed it might be context heavy, unlike RepoMapper which an set the Context limit? I'll have to see...

[-]

Dundell@reddit (OP)

Nvm context7 was very easy to setup with no need further past the MCP marketplace within RooCode. Just load it, always allow, and asked my gpt-oss 120b guy to use it to lookup some things like my newspaper4k implementation. Shows 2,000 context limit set and some general information pulled about uses. So that seems useful, probably more useful for npm's.

[-]

Niku_Kyu@reddit

gpt-oss tool call support is 🤢 right now.

[-]

Dundell@reddit (OP)

I've been busy working on the MCP server from "[pdavis68/RepoMapper](https://github.com/pdavis68/RepoMapper)" , forking it and reworking it a bit for my local files. So far on 4 different project/repos I have it works fine for the MCP tool calling Of course I do have my server running the configs: /mnt/sda/model/llama.cpp/build/bin/llama-server -m /mnt/sda/llama.cpp/models/gpt-oss/gpt-oss-120b-UD-Q4\_K\_XL-00001-of-00002.gguf --ctx-size 131072 --flash-attn --n-gpu-layers 48 --tensor-split 4,6,6,6,6,8 --host [0.0.0.0](http://0.0.0.0) \--port 8000 --api-key YOUR\_API\_KEY\_HERE -a GPT-OSS-120B-K-XL-Q4 --temp 1.0 --top-p 1.0 --top-k 100 --threads 28 --jinja --chat-template-kwargs '{"reasoning\_effort": "high"}' --chat-template-file /mnt/sda/llama.cpp/models/gpt-oss/gpt-oss.jinja --grammar-file /mnt/sda/llama.cpp/models/gpt-oss/cline.gbnf Which I believe the last part the grammar-file helps with tool calling in RooCode. Least I rarely see any failures.