new samples for mormot2, any rules?

flydev · 2024-10-26 09:30:58

@ab is there any rules to follow before making PRs about pushing some samples dedicated for mormot2?

I finished to write a vanilla js lib which support mormot custom auth scheme that will be available on github and i would like to provide samples projects about it.

I have other projects as well, like client/server websockets, an mvc dashboard, etc. I thought they might be useful to other developers.

zen010101 · 2024-10-28 08:37:39

It sounds great!!

ab · 2024-10-28 17:00:40

I guess that if you put the samples in the ex/thirdpartydemos folder, in a sub-folder named by the writer, we could integrate them.

A README file would be great too, to explain what each sample is about.

mpv · 2024-10-28 20:41:18

Perhaps a good way is to put the third-party samples in your own repository and link to it in ex/README.md (many well-known projects do this, calling it a curated list, etc.). In this case, the owner of the third-party repository can change their samples at their own discretion, and the only PR for the mORMot repository is changes to ex/README.md

flydev · 2024-10-28 20:49:13

A README file would be great too, to explain what each sample is about

Sure, thanks.

I like the idea @mpv

Last edited by flydev (2024-10-28 20:51:08)

edwinsn · 2024-10-29 02:03:06

@flydev, that sounds very interesting, couldn't wait to see what you has to offer!

I agree with MPV, if you have the plan to continually improve it, you might want to make it a standalone repo on github, so that even the others can make PRs.

flydev · 2026-02-08 20:10:15

I still haven't made a PR yet, but a public repo is finally live where some "extensions" as i'm calling them will be added. The extensions listed as upcoming will be pushed this week after adding a readme before pushing.

For now, you can find a pure mormot2 MCP implementation that works on windows and linux and can be compiled with Delphi (at least 12.3) and FPC. Just start with the tests then the demos.

@zen010101 I hope you will test this one!

Last edited by flydev (2026-02-08 20:14:56)

zen010101 · 2026-02-17 07:14:19

Thanks a lot @flydev! Really appreciate you sharing this mormot2 MCP implementation and the extensions repo — it looks great!
Just to give you a heads-up: I’m currently buried deep in a production project,so I probably won’t have time to test it properly until after the end of March, unless the project wraps up early unexpectedly.
Once I’m free, I’ll definitely give it a spin and provide feedback! ?

flydev · 2026-02-17 08:17:00

I'm also on a rush, last sprint! let's focus. On end of march you then find an experimentation that give to agents access to the computer resources, like opening softwares, switching between window with keyboard/mouse input. I already made a test where the agent open mspaint and start drawing, it was funny to watch

I see I didn't put any link to the repo on my last message as I didn't made a link on mormot ex yet. Added winsparkle wrapper, and a TOTP implementation will be added this week, quite useful.

For people interested, repo is there: https://github.com/flydev-fr/mormot2-extensions

igors233 · Today 04:10:50

I think MCP is a new thing, atleast is for me. What is the usecase and is that a local AI model or connects to external AIs?

flydev · Today 10:46:40

MCP servers are cool. MCP is not an AI model. It's more a bridge that gives the model access to tools, local or remote (and i'm sure you already heard about Skills too).

For example, with an MCP you can expose a tool so the model can fetch specific data from your REST service (for example using "orm.RetrieveDocVariantArray()"), and then analyze this data with a dedicated Skill.

If we take the mvc-blog sample:

- the MCP gives the model a tool to return a list of posts based on your own criteria,
- the Skill lets the model act like an analytics expert and produce a summary.

We could then automate the SEO and posting of our blog - you can see how the model could update or rewrite articles with this kind of flow:

1. MCP: get analytics data from your Google SearchConsole
2. Skill: analyze low-ranked articles
3. MCP: fetch the articles from your system
4. Skill: analyze the content and think about improvements
5. MCP: update the post in the database

So the Skill handles the thinking and SEO logic, and the MCP handles the access to real data and real actions.

Another example: if your MCP exposes tools like "FindWindow()", "MoveMouse()", "SendKey()", etc. then the model could really interact with your computer.

So basically:

Skill = how the model thinks (expert logic).
MCP = what the model can do (actions).

Most of the time you need both: a skill for reasoning, and an MCP to connect and act in the real world. Possibilities are endless.

flydev · Today 11:24:25

A real example on this page, there is a public tool that allows parking users to retrieve their payment receipt using their vehicle license plate and the parking date.

If the given plate is not found exactly, the system can return a list of similar plates based on our own logic. Then AI use:
- an MCP tool to query and retrieve this candidate set,
- a dedicated Skill to analyze the results and determine the most probable correct receipt.

Another real world example I'm currently experimenting is AI support through a SIP server:
- realtime model answer support calls by voice.
- the model get access to a particular parking resources based on enabled modules: could be the existing license plates stored on local computer (from where the call happen), an issue with their member's badge, etc.

In this case:
- SIP server model is connected to the MCP and get dedicated backend services (plates, badge-numbers, some hardware statuses, etc.)
- Skill handles the reasoning, understands the caller request, and generates the proper spoken response or trigger a notifications so an human get involved.

Again, MCP handles the data access, and the Skill handles the intelligent decision.

realtime models is a version of GPT - capable of responding to audio and text inputs in realtime over WebRTC, WebSocket, or SIP connections.

As I said, possibilities are endless

dcoun · Today 12:10:34

That is excellent, it can help mormot2 to gain more appearance and users

igors233 · Today 13:10:58

Thanks for this, quite informative. So MCP is kind of a standard that AI agents already knows and you just add support for what you need from your ERP side.

Are Skills agents from Claude?

In your video there's a part of understanding what's written and performing task, is that Skills part?

I was thinking of adding some sort of that to software, respond to written request (I think that's called NLP) and perform something.
Do you know of a system that could be run locally, without paid subscription and create good results?

flydev · Today 15:53:08

igors233 wrote:

So MCP is kind of a standard that AI agents already knows and you just add support for what you need from your ERP side.

Exactly.

In the video there was not skill involved, I just connected the MCP (from the demos folder) and asked the agent to return my ip available mcp tools ("public_ip" that return ip and time, see there).

But a skill could have been set to tell the agent what to use and what to call when asked to return a public ip. You can read on @openai codex repo or claude cli, etc to see how they write skills there. I suggest you to read an introduction on what are skills, maybe on anthropic or openai blog, doesn't matter.

igors233 wrote:

Do you know of a system that could be run locally, without paid subscription and create good results?

I will not get into hardware things, you know that if you run things on CPU it will be slow as f*** - that's said, you have a lot of available tools like ollama, lama.cpp, lmstudio.

To get started I recommend LM Studio:
- Really easy to use. It's easy to fetch, load and experiment different open source models - you can even literally paste the url of a model from huggingface.co into the app, it will download and be ready to use. You can also start an api server from the app to interact with.

I suggest you start with following models: Mistral, Llama 3, or Qwen and for more models look there, @TheBloke's published models

- fun but experimental, stay with Pascal: https://github.com/BeRo1985/pasllm

mORMot Open Source

#1 2024-10-26 09:30:58

new samples for mormot2, any rules?

#2 2024-10-28 08:37:39

Re: new samples for mormot2, any rules?

#3 2024-10-28 17:00:40

Re: new samples for mormot2, any rules?

#4 2024-10-28 20:41:18

Re: new samples for mormot2, any rules?

#5 2024-10-28 20:49:13

Re: new samples for mormot2, any rules?

#6 2024-10-29 02:03:06

Re: new samples for mormot2, any rules?

#7 2026-02-08 20:10:15

Re: new samples for mormot2, any rules?

#8 2026-02-17 07:14:19

Re: new samples for mormot2, any rules?

#9 2026-02-17 08:17:00

Re: new samples for mormot2, any rules?

#10 Today 04:10:50

Re: new samples for mormot2, any rules?

#11 Today 10:46:40

Re: new samples for mormot2, any rules?

#12 Today 11:24:25

Re: new samples for mormot2, any rules?

#13 Today 12:10:34

Re: new samples for mormot2, any rules?

#14 Today 13:10:58

Re: new samples for mormot2, any rules?

#15 Today 15:53:08

Re: new samples for mormot2, any rules?

Board footer