Hi,

Another option is to distribute MCP server publishing and SKILL/Agents
marketplaces. Or, as JBO mentioned, use CLIs and remote APIs.

MCP servers are often (more often than not) not actual remote servers that
need to run "somewhere else". Many run locally, outside the agent's
sandbox, and expose the MCP server on a local stdin/connection, shielding
the MCP logic from the agent. Their server functionality is merely the API
they expose to local agents.

I think there are two broad models we can discuss: centralized
(framework-based) and distributed (standards-based). Both can coexist
within the ASF. I can very easily **everything** we dp routed by ASF
`llmao` and "tight controlled" by the infrastructure / tooling - but also I
can very easily imagine that ASF contributors and maintainers use their own
LLM /  Subscriptions and local servers, connecting to ASF APIs - this is
IMHO more scalable and more "Apache Way" approach. Both models have pros
and cons, but with the standards published in recent months, the barrier to
using the distributed approach has dramatically lowered.

I personally strongly favor a distributed approach over a centralized one
for MCP servers, SKILLs, and Agent definitions. This allows each component
to be maintained and published separately by different PMCs directly in
their GitHub repos. It is sufficient for their code to be published
somewhere for the MCP server to be installed and run locally. For example,
the MCP servers for Ponymail and Apache Project MCPs in Comdev —which we
use in Magpie—are both simply installed and updated from the Comdev GitHub
repo.

Standards for all of this were established in recent months (MCP 2.0 and
Agents Marketplace standards established in July/August).

We recently converted Magpie to be installed from a distributed marketplace
in preparation for Magpie's 0.2.0 (or maybe even 1.0.0) release. ) release.
I am very happy with the test results; the open standards work smoothly
across many agent harnesses and are already quite powerful.

For example, I used (rather than copying as I was doing so far) the threat
model skills from Alpha-Omega to handle threat modeling for our scan
preparation. I plan to make this accessible for all PMCs and non-ASF
projects—more details will be shared in Glasgow, where we plan to run
workshops addressing PMC needs around managing security issue volume
(including those from our scans).

Moving forward, I plan to pursue what we previously discussed on the Magpie
devlist with Comdev and Incubator PMCs: distributing skill families back to
Comdev and Incubator so they can be managed and exposed directly by those
PMCs via distributed marketplaces, rather than hosting them in the Magpie
repository. There are few benefits to centralizing them in one place; it's
much better to have them managed by the respective PMCs and simply "link"
to them.

Similarly, Legal could distribute their licensing, MCP, and SKILL
marketplace - and we could link to it from Magpie and expose the
"licencing" family of skills.

This approach - I think - supports better scaling and distributes
governance in the true "Apache Way," giving each PMC autonomy over
implementation languages, scope, and publishing methods using open
standards.

Best,
Jarek

On Tue, Sep 8, 2026 at 11:45 AM Jean-Baptiste Onofré <[email protected]>
wrote:

> That's a valid point: magpie is definitely worth a try on podlings.
>
> Regards
> JB
>
> On Tue, Sep 8, 2026 at 8:12 AM Justin Mclean <[email protected]>
> wrote:
> >
> > HI,
> >
> > On the AI side, Apache Magpie has a release skill that will find issues
> in release candidates for you. I suggest that prodlings try this out. Also,
> the Policy MCP can be used to check releases. Having teh Policy MCP make it
> consume a lot fewer tokens, faster, and up to date if the policy does
> change.
> >
> > Kind Regards,
> > Justin
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: [email protected]
> > For additional commands, e-mail: [email protected]
> >
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: [email protected]
> For additional commands, e-mail: [email protected]
>
>

Reply via email to