What's your favorite local MCP server?

[-]

vamps594@reddit

Sentry and Grafana MCP are nice and save a lot of time.

[-]

Glittering_Focus1538@reddit (OP)

Can you explain the usecase for both?

[-]

I mainly use Sentry MCP at work. It saves me a lot of copy-pasting between Sentry and my IDE (for example: "fix Sentry issue XXXX-YYYY", "analyze slowdowns on query X", etc.). It also helps me monitor and categorize errors and logs (critical, normal, and so on).

I use Grafana MCP mainly to create and update dashboards and set up alerts. Most of the time I already know what I want, but creating all the charts for each query and adjusting every panel manually takes a lot of time. Usually, I can generate about 90% of the dashboard automatically and only need to make around 10% of the adjustments by hand.

[-]

Fit_Squash6874@reddit

Depends on what I want to do. I mostly build my own mcp.

[-]

MaxKruse96@reddit

Serena, although getting LLMs to use it is flakey

[-]

Glittering_Focus1538@reddit (OP)

doesnt seem to have a mode for kiro unfortunately, looks cool tho.

[-]

New_Zone5490@reddit

serena can be set up to work on kiro

[-]

Glittering_Focus1538@reddit (OP)

i didn't see a preset for it when I set it up :/

[-]

Left-Orange2267@reddit

You don't need a preset, they are all roughly the same anyway. Use the --context ide and you'll be fine.

I am one of the creators of Serena, so the suggestion is sound ;)

[-]

New_Zone5490@reddit

thank you for creating this valuable tool

[-]

Left-Orange2267@reddit

Thanks for the kind words!

[-]

Glittering_Focus1538@reddit (OP)

Thanks boss, but consider adding it, I saw claude code and codex and others and went, welp maybe this isn't for me.

[-]

New_Zone5490@reddit

add this to ~/.kiro/settings/mcp.json:

{
  "mcpServers": {
    "serena": {
      "command": "serena",
      "args": [
        "start-mcp-server",
        "--context=ide"
      ],
      "disabled": false,
      "autoApprove": [
        "activate_project",
        "find_symbol",
        "search_for_pattern",
        "get_symbols_overview",
        "read_memory"
      ]
    }
  }
}

the "autoApprove" part is optional, but they will get added anyway once you click "approve" when agent first attempts to use them. so you could just copy paste that part too

[-]

patricious@reddit

Upvote for Serena, to make the harness lookup symbols you need to add that instruction to your local AGENTS.md

[-]

i_like_brutalism@reddit

i always build mcp servers based on the projects i am doing. if i need a certain api or tool to be available, i quickly whip one up.

recently i needed to interface with a serial device and also built a mcp server for that. i know the trend goes towards just cli, but if the shell commands are super advanced and require reset commands inbetween, i feel more comfortable hardcoding them inside a mcp server.

[-]

Glittering_Focus1538@reddit (OP)

Fair enough, so 1 off mcp's are ur thing now?

[-]

i_like_brutalism@reddit

yeah, mostly. i do a lot of cybersecurity stuff, so e.g. GhidraMCP (although sadly not that great with smaller local models) and a few more run off the mill servers are useful too.

but custom tooling (incl mcp servers) really improves the workflow for bigger projects! for me thats one of the major advances since llms became a thing. i never really think if ifs worth implementing a small tool, since it can be done so much faster now

[-]

floppypancakes4u@reddit

I was wondering if they was a ghidramcp. Publicly avaliable i hope

[-]

i_like_brutalism@reddit

yes, https://github.com/bethington/ghidra-mcp

but this implementation requires a lot of context to work well

[-]

taeper@reddit

Have you tried REVA mcp? https://github.com/cyberkaida/reverse-engineering-assistant worked well for me although i don't do much reversing

[-]

i_like_brutalism@reddit

will try it out!

[-]

floppypancakes4u@reddit

context as in information, or the context window?

[-]

i_like_brutalism@reddit

apologies, should have been more clear. im talking about the context window!

[-]

floppypancakes4u@reddit

Gotcha. Yeah I can imagine more is better here. What do you recommend with your experience so far?

[-]

i_like_brutalism@reddit

completely vibes based experience: i was able to run it with smaller local stuff, and the intelligence was always okay-ish. for small tasks and understanding of simple projects that might suffice. but as soons as you have some pointer tables, or weird branches, local models start showing their weaknesses

[-]

Glittering_Focus1538@reddit (OP)

You need a model qwen 27b class or higher, would probably stick to SOTA models if you want it to run well.

[-]

old-mike@reddit

Might sound silly, but I'm actually using Hermes as a personal assistant. I'm just starting out, but right now I have the LinkedIn MCP connected to Hermes Agent to help me polish and update my profile.

Hermes fetches a structured list of my notifications and connection requests, presents them as a clean action-item list, and then I just check off what I want it to execute. On top of that, I set up a cron job every 30 minutes to ping the MCP and alert me if there’s any reply from job applications I've submitted. Still fine-tuning that last part, but the automation loop is active.

[-]

Glittering_Focus1538@reddit (OP)

I would use hermes if I could run qwen 3.6 constantly, but it takes up too much gpu for me to game at all with hermes running and it cant do much with gemma 4, so unfortunately I can't copy this method.

[-]

cpt_justice@reddit

llama-swap, maybe? When gaming, unload the model. When done gaming, load the model. llama-swap keeps on running 24/7.

[-]

giveen@reddit

llama.cpp has a built in router mode now and just put in a 600sec timeout

[-]

cpt_justice@reddit

That's what I was using prior to llama-swap. What got me to switch was llama-swap will... well... swap models out at will and a way to pick which model to unload when loading a different one.

[-]

Comfortable_Ebb7015@reddit

Well, you can pause it while gaming, no?

[-]

Glittering_Focus1538@reddit (OP)

kinda defeats the whole purpose of a 24/7 assistant no? and I game a lot T_T

[-]

Solary_Kryptic@reddit

You could set up the server to kill when you load up a game and automatically load up again once the game closes. I don't see why you would need the assistant while playing

[-]

AnticitizenPrime@reddit

I understand why. One of the ways I use Hermes is to have it do research on stuff in this background while I do other stuff, and have it ping my phone when it's ready. So I can understand wanting to game while the AI chugs away working for you. Unfortunately it's just one of those situations where you either need a rig capable of both, a separate rig for games, a console, whatever.

[-]

AnticitizenPrime@reddit

Have you ever considered the wide and wonderful world of retro gaming? lol

[-]

roosterfareye@reddit

Just whack a second card in your secondary slot. Profit.... Well, you need to these days lol!

[-]

Glittering_Focus1538@reddit (OP)

board doesnt have one sadly.

[-]

CountVonTroll@reddit

I'm still in the process of setting my local assistant up, but I've configured the default local model to get called via LiteLLM router, because it lets me configure a fallback.

[-]

Guinness@reddit

cloakbrowser

[-]

Glittering_Focus1538@reddit (OP)

have you tested it against crawl4ai?

[-]

Sevealin_@reddit

Frigate MCP for my security cameras, awesome for config changes/suggestions - https://github.com/dedsxc/mcp-frigate

Unofficial Home Assistant MCP for creating automations and troubleshooting - https://github.com/homeassistant-ai/ha-mcp

n8n-mcp for creating n8n automations (that hit my llama.cpp) on the fly and troubleshooting broken ones - https://github.com/czlonkowski/n8n-mcp

[-]

Glittering_Focus1538@reddit (OP)

Sick

[-]

Rooneybuk@reddit

My big go to are n8n, Grafana and playwright

[-]

Glittering_Focus1538@reddit (OP)

What does Grafana do? mcp code graph indexer?

[-]

9gxa05s8fa8sh@reddit

recently I'm testing:

for memory: https://github.com/rtk-ai/icm

for documentation/library/research:

[-]

Endurance_Beast@reddit

Chrome-Devtools. Sometimes I needy local model to do some manual labor work and devtools is great for that.

[-]

AnticitizenPrime@reddit

What sort of thing are you having it do?

[-]

Endurance_Beast@reddit

Filling up forms for services where I don't have api keys for, chatting with another SOTA via web gui to discuss a solution to a problem, those sorts of things.

We ran into a problem where a client needed to migrate their data from an online OKR platform last month, they had an api key with endpoint to the OKR module only. The users, projects, strategy and performance lacked. So chrome-devtools came to the rescue.

I created an agent that will scrape and fill up the data in the new format instead of writing a mapping script. Qwen3.6 27B finished that alone with this MCP.

[-]

Glittering_Focus1538@reddit (OP)

OOO good to know, thank you!

[-]

Primary-Research-747@reddit

same boat, run 4 daily too. mine:
chrome-devtools map - agentic browser control without paying for browserless. probably highest leverage one I use is supabase map - read/write directly into project DBs from Claude Code, no copy - paste.
claude-in-chrome different from chrome-devtools, drives my actual logged-in browser for stuff that needs auth and a custom one for content ops.
cut last month: GitHub mcp (gh cli just better imo), filesystem mcp (claude code's built in tools cover this), memory mcp I was testing.
rule of thumb I landed on - mcp that bridge claude to systems it can't otherwise touch - keep. mcps that wrap things claude already does well = net negative

[-]

OpenClawInstall@reddit

My boring answer: the best local MCP servers are usually the ones you write for one specific workflow. Big generic tool packs look impressive, but they make models worse at choosing the right action.

The local MCPs I keep reaching for are basically:

filesystem, but scoped hard to the project
sqlite or a tiny task/state store
browser/devtools when the job is UI verification
a custom wrapper around one annoying internal command

The custom wrapper matters because it turns five fragile shell steps into one named tool with typed inputs and predictable output. That is way easier for an agent to use safely than giving it a full terminal and hoping it remembers the cleanup step.

My trimming rule is simple: if a tool has not saved a real run in the last week, remove it. A smaller MCP surface with better descriptions beats a giant menu almost every time.

[-]

ag789@reddit

well, pretty much a novice, but that I make an MCP server to simply let it run some shell commands e.g. ls, cat, echo, grep, date, etc lol
it turns out this is pretty practical, e.g. you can ask what is the date today and get a correct answer with even smaller LLM.

[-]

DiscipleofDeceit666@reddit

The one that I built myself lmfao 1 single MCP endpoint that gives cloud AI the ability to ask my GPU AI questions about my file base. It’s part of a project where my GPU subsidizes cloud code writing operations with very limited vram. But ofc, any open router endpoint will work.

www.GitHub.com/minerest/leanloop

[-]

i_like_brutalism@reddit

test

[-]

dev_dan_2@reddit

Currently building my custom one (mainly for coding, assuming the reasoning is done by me or a bigger LLM, and then my small LLM shall use my MCP server to reliably do the small subtasks.)

My main features I want to have:

everything that happens happens because the MCP server executes it (for example, it never calls bash, but instead interpretes the command and then uses the bash crate.)
everything is forbidden unless confirmed by me or explicitely allowed by being in an allowlist
MCP should do the heavy lifting, the LLM should do as little as possible

I do not see this managing my social media stuff (which I rarely use anyway …\^_\^'), but I want to be able to do local development and I would rather have my tooling do exactly what I want and need. Thanks to LLM that is easier than ever, too! :D

[-]

JRufer@reddit

I've been working on a voice transcription app called VoxCtr. It uses a targeting system to route your text to different services as needed. I wanted a way for your services to talk back to you so it includes a small TTS MCP server that lets your Hermes Agent, or any other process, talk back to you in response.

[-]

UnWiseSageVibe@reddit

firecrawl has been super useful and nice.

[-]

Glittering_Focus1538@reddit (OP)

have u compared it to crawl4ai?

[-]

New_Zone5490@reddit

serena

[-]

Glittering_Focus1538@reddit (OP)

Cant use it with kiro unfortunately, seems cool tho

[-]

Parzival_3110@reddit

For local MCP stuff, the browser one is the category I keep coming back to.

A lot of agent workflows are fine with files and shell, but the moment the task touches a logged in site, docs behind auth, extension state, or a real app preview, plain scraping stops being enough. You want scoped tabs, DOM reads, screenshots, action logs, and a clear boundary around side effects.

Bias disclosed since I am building one: FSB is the browser MCP I use for that shape, especially with Claude Code and Codex. It gives agents a real Chrome tab without making every task a remote browser service.

https://github.com/LakshmanTurlapati/FSB

[-]

Glittering_Focus1538@reddit (OP)

how does this compare to something like Crawl4ai?

[-]

henk717@reddit

openzim-mcp combined with ddgs, and of course I use KoboldCpp's mcp bridge to bundle it all (Which is not starlette based). Openzim lets me load it with offline data I can fact check which I did for emergency situations if the internet ever goes down, mine is loaded with wikihow so I can get reliable howto's (I verify it with a zim viewer when it finds stuff). And then ddgs as a good search engine since the one inside koboldcpp is minimal by comparison.

[-]

sahanpk@reddit

filesystem + sqlite are the boring ones i keep coming back to. custom one-off MCPs beat giant tool packs for real projects.

[-]

Steve_Streza@reddit

Context7 is pretty good at getting information about libraries.

[-]

Glittering_Focus1538@reddit (OP)

Already using the free tier T_T, got that, ladybugdb, local rag and desktop commander goin.

[-]

PixelSage-001@reddit

The standard filesystem and memory MCP servers are definitely the most useful day-to-day. However, the real challenge isn't just having the server, it's how the agent decides to invoke the tools without getting stuck in infinite loops.

I ended up setting up my agent workflows using Runable as the orchestration layer. It lets me define strict boundaries on when and how local MCP tools are triggered, so the model doesn't go off-track or bloat the context window with repetitive file reads.