feat(mcp): improve neighbourhood publishing with auto-cloning

[data-bot-coasys]

Problem

AI agents (including me when building the SoA prototype today) bypass MCP and fall back to GraphQL because the neighbourhood publishing flow requires too much low-level knowledge:

• Manually clone link languages
• Understand language addresses and templates
• Know the exact sequence: clone → publish → use address

This defeats the purpose of MCP — it should be the only interface agents need.

Solution

Three improvements:

1. list_link_language_templates — new tool

Returns available P2P sync engine templates (currently just Holochain perspective-diff-sync). Agents pick from this list instead of manually looking up addresses in config or GraphQL.

{
  "templates": ["QmzSYwdn..."],
  "count": 1,
  "hint": "Pass one of these as link_language_template when publishing"
}

2. Auto-clone in neighbourhood_publish_from_perspective

Agents now pass:

• perspective_uuid — what to share
• link_language_template — which sync engine (from list_link_language_templates)
• name — human-readable neighbourhood name

The tool handles:

• Cloning the template with unique ID
• Publishing the clone via the language language
• Using the cloned instance for neighbourhood sync

No manual language management. One-step workflow.

3. Strip implementation details from descriptions

Tool descriptions now explain what ("publish a perspective as a neighbourhood") not how ("requires a perspective-diff-sync language address QmzSYwdn...").

Agents work at a higher level of abstraction.

Before/After

Before:

// Agent needs to know language cloning workflow
template = "QmzSYwdn..." // where did this come from?
cloned = language_apply_template_and_publish(template, '{"name": "My NH"}')
neighbourhood_publish_from_perspective(perspective_id, cloned.address)

After:

templates = list_link_language_templates()
neighbourhood_publish_from_perspective(
  perspective_id, 
  templates.templates[0], 
  "My Neighbourhood"
)

Testing

Verified with:

• cargo check — compiles clean
• Manual GraphQL/MCP comparison — same capabilities, simpler interface

Next: test with actual neighbourhood creation (will do with proper SoA memory perspective).

Related

Part of making MCP the canonical AI-agent interface. Closes the loop on why I fell back to GraphQL when building the SoA prototype — this would have prevented that.

Summary by CodeRabbit

• New Features

  • Can list available link-language templates for neighbourhood creation.
  • Added a language metadata lookup tool to inspect installed languages.

• Improvements

  • Neighbourhood publish now auto-clones templates to create unique link-language instances.
  • Publish responses now include the cloned link-language address and neighbourhood name; publish/join flows use the cloned instance.

• Documentation

  • Expanded docs for language & neighbourhood tools, publishing and joining workflows.