🎉 Announcing the first Community Health Toolkit MCP server!

AI
1

We are excited to announce that the first Community Health Toolkit MCP server is now available!

This integration lets AI applications connect directly with CHT knowledge, including updated CHT documentation, GitHub issues, code examples, community knowledge, and forum posts.

:backhand_index_pointing_right: Try it out by following the dedicated tutorial.

A huge shoutout to @hareet and @sugat for making this happen! :raising_hands:

4 Likes
2

Wow @andra have we open sourced it? wanted to take a look at architecture

3

You can find a high-level architecture diagram of the CHT Docs MCP on the docs site. In a nutshell, it uses Kapa AI integration to retrieve relevant CHT knowledge.

@hareet can help answer any detailed questions you might have.

1 Like
4

@shashi_sah - We’re Medic - of course it’s open source :smiling_face_with_sunglasses:

The CHT Docs MCP Server is a key component of the CHT Agent multi-agent system for automating CHT development tasks.

Specifically the documentation-search-agent.ts file likely has what you’re looking for.

Oop - Andra just posted almost the same answer :sweat_smile:

1 Like
5

I think @andra won in the race condition :joy:

2 Likes
6

@Romuald_Dansou - Could you share some of your experience with Claude and generating your CHT Specialist skill? Specifically, I think you had found that when trying to use the CHT MCP you were running out of contexts so you explored using skills maybe?

I think real world experience is super helpful if you’re able to share!

7

Yeah I agree using skill with Claude is a good idea. Maybe having a directory with different sets of skill like debugging cht, exploring cht etc might be helpful with Claude?

Someone could use plugin in Claude to use those skills locally also, we can also have those skills integrated with mcp server

8

Excellent point on the Claude skills @shashi_sah. @hareet has recently opened a PR for documentation creation and documentation review Claude skills.

We had discussed in the past ideas for developing debugging skills for CouchDB, but there are, of course, many other skills that could be helpful within the CHT. More to come soon - stay tuned!

9

Great suggestions everyone!

We can review these points in our first meeting.

To clarify the design choices: the CHT-Agent targets non-technical users and must be model-agnostic. The points below don’t make Skills the best choice for our documentation access layer.

Some benefits of MCP over Claude Skills for our use case:

  • Managed RAG: Kapa AI is handling all the heavy lifting for us. They handle continuous indexing, embedding optimization, and response quality. With Skills, every user would effectively build their own local RAG, duplicating infrastructure and missing the benefits of a centralized, maintained knowledge base.
    • Skills teach AI Agents how to perform tasks (procedures, workflows). MCP provides connectivity to external data and services. We don’t want to replicate and maintain the same feature the external service is providing.
  • Multi-Agent design: In our architecture, the documentation agent only handles docs queries. Context window concerns from MCP tool usage don’t apply when an agent has a single, focused responsibility.
  • Model portability: The CHT-Docs MCP Server works with any model. Skills are being ported to other models as of this month, but MCP is more supported / standardized.

Skills are great and valuable for technical power users customizing their personal AI workflows. We want all your Skills and shared use cases! Skills will be valuable for some of the other agents and their tasks. Organizationally, we need to be intentional about preparing CHT for agentic infrastructure for community members who won’t configure prompts or maintain embeddings.

@Romuald_Dansou @shashi_sah The end goal is someone like Loukman / Program Director being able to request a feature for cht-core through the multi-agent system, without needing to understand the underlying stuff. The MCP docs server is a foundational piece that lets us deliver that while keeping everything model-agnostic, and using updated information via managed-services. We want to abstract that further and let other community members run the cht-agent locally, in their own datacenters, using their own models, and having the ability to query each agent independently.

A good summary and comparison chart by Claude:

Skills = curated expertise for specific tasks (writing forms, debugging tasks.js)
MCP Docs Server = searchable knowledge base for any CHT question

For the cht-agent’s goal of serving non-technical community members like Loukman, we need the breadth of semantic search. But Skills like cht-specialist could absolutely power a specialized configuration agent within the multi-agent system.

Screenshot 2026-01-23 085705
Screenshot 2026-01-23 085705468Ă—433 30 KB

3 Likes
10

Thanks @hareet for detailed information. Just couple of questions I wanted to ask

  1. Are we assuming Kapa AI as a long-term dependency, or we gonna build our own RAG and have vector db? also does kapa ai index everything again or just the latest changes?
  2. Do we have any strict boundary between different agents? For example will the docs mcp agent only answer queries related to docs or will it have connector for other mcp as well?
11

Great questions!

As of currently, we would prefer not to maintain our own. We will have to revisit this periodically to see if that decision is still valid. We named this MCP server “CHT Docs MCP Server” to help identify it may not always be Kapa AI, it could be a different vendor.

Do we have any strict boundary between different agents? For example will the docs mcp agent only answer queries related to docs or will it have connector for other mcp as well?

Let’s move this discussion to the #multi-agent-system-squad Slack channel. To close the loop here: We will start with narrow boundaries and lightweight connectors, but if you find something helpful for an agent and their tasks: let’s pack the configuration in there!

1 Like