Stellar Ecosystem Scout

Turn a sector or topic ("RWA tokenization", "wallets", "DeFi", "cross-border payments") into a structured landscape: the projects that fill it, how they group by category and region, the controlled tags that describe them, and the editorial / news / talk / research content that explains the space — every content claim tied to a source url. This skill runs on LumenLoop's free, read-only Stellar ecosystem MCP at https://mcp.lumenloop.com (the public project directory, SCF data, and content library). It maps a space; it does not deep-dive one team or pick one dependency.

When to use this skill

• The user asks "what projects exist for X on Stellar" or "map / survey the X space".
• You need a landscape of a sector grouped by category and region before going deeper.
• You want to grow a cluster from one known project to its semantic neighbors.
• You need the narrative around a sector — articles, talks, research — with cited urls.
• You want to spot gaps / white space in a category before building or investing.
• The user names a fuzzy theme ("regenerative finance", "gaming") and wants the serious players plus the supporting content surfaced.

Related skills

• Set up / troubleshoot the connector and see all 18 tools → ../lumenloop-mcp-connect/SKILL.md.
• Deep-dive ONE project you surfaced here (funding, talks, comparables) → ../stellar-project-dossier/SKILL.md.
• Pick ONE dependency to integrate from a landscape (wallet/oracle/anchor/…) → ../stellar-integration-finder/SKILL.md.
• Turn the landscape into a build plan (primitives, prior art, stack) → ../stellar-builder-quickstart/SKILL.md.
• For actually BUILDING on Stellar (Soroban contracts, JS stellar-sdk / Freighter / Stellar Wallets Kit, classic assets + trustlines + SAC, Stellar RPC / Horizon data, SEP/CAP standards, Reflector oracle), see the Stellar developer docs at https://developers.stellar.org.

The scouting recipe

Work in five passes. Each pass is one or two tool calls; stop when the cluster stops growing and the narrative is covered.

Pass 1 — Frame the taxonomy

Pull the controlled vocabularies first so you bucket everything consistently and can name gaps in the project's own language.

get_categories()
get_regions()
get_project_tags_vocabulary()

get_categories() returns the five live buckets — Applications, Developer Tooling, Education & Community, Financial Protocols, Infrastructure & Services. get_regions() returns distinct based_in / operating_region values for the geo view. get_project_tags_vocabulary() returns the controlled PROJECT tags (e.g. "Wallet Infra", "Cross-Border Payments", "SDK", "RWA") — these are the precise labels to slice the landscape by, and the best source of SHORT seed terms for Pass 2. (Note: get_tags_vocabulary() returns CONTENT tags for articles/AV, not project tags — don't confuse the two.)

Pass 2 — Seed by keyword (mind the ILIKE gotcha)

search_directory(query, limit?) is a keyword / ILIKE match over project title, description, and other_names. SHORT terms win. A long natural-language phrase matches nothing.

search_directory(query="real world assets tokenization")   # → 0 results (too long)
search_directory(query="RWA")                               # → real hits
search_directory(query="tokenization")                      # → real hits
search_directory(query="asset")                             # → real hits

Fire several short, orthogonal queries to cover synonyms (for RWA: RWA, tokenization, asset, stablecoin, real estate). Results come back as one JSON object {count, projects}; each project row is {slug, title, description, category, tags, website}. Collect the slugs — these are your seeds. If a query returns 0, the response includes a recovery hint — shorten the term or pivot to the semantic tools below.

Pass 3 — Grow the cluster semantically

Pick the most central seed slug and expand to its semantic neighbors. This beats tag/category matching for finding adjacent players you didn't have a keyword for.

find_similar_projects_semantic(slug="<seed-slug>", limit=10)

It returns projects similar to the seed by embedding distance. Run it on two or three different seeds and union the results to triangulate the cluster boundary; the overlap is the dense core of the sector, the long tail is the frontier.

Do NOT use get_related_projects for project→project — that one maps a CONTENT item to the projects it mentions (Pass 4). For project→project, always use find_similar_projects_semantic.

Pass 4 — Add narrative + cite sources

Layer the story on top of the project list with semantic content search. This is also the rescue path when search_directory returned 0 for a long phrase.

search_content_semantic(query="real world asset tokenization on Stellar",
                        types=["articles"], limit=8)

types ⊆ articles | av | events | research (omit for a broad sweep; narrow for speed). limit defaults to 5 per content type. It returns per-type arrays of {id, title, url, domain, publishing_date, summary, slug, similarity} — pass response_format="detailed" if you also want each hit's long_summary. These are LumenLoop's AI-generated summaries plus the source url — cite the url, never quote them as full article text. A query like the one above surfaces real explainers such as "tokenization-basics".

For any salient article, map it back to the projects it discusses:

get_related_projects(content_id=<id-from-content-result>, content_type="article")

content_type ∈ article | av | event | research (plural/alias ok); the source must be published. This often surfaces projects you missed in Pass 2/3 because the article named them but the directory keyword didn't match. To pull direct quotes from talks or podcasts about the sector, use find_av_passages(query="<sector>").

Pass 5 — Detail each project (slim shape)

For every project that makes the landscape, pull the writer-friendly slim row:

get_project(slug="<slug>", compact=true)

compact=true returns ~500 chars — slug, title, short description, category, tags, twitter_handles, website, other_names — exactly what a landscape table needs. Omit compact (default full row) only when you genuinely need the long record; for a sector map the compact shape keeps it tight.

Worked example — mapping RWA / tokenization

# Pass 1 — frame
get_categories()                  # → Applications, Developer Tooling, Education &
                                  #   Community, Financial Protocols, Infrastructure & Services
get_project_tags_vocabulary()     # → includes "RWA", "Stablecoins", "Cross-Border Payments"…

# Pass 2 — seed (short terms only)
search_directory("real world assets tokenization")  # → 0 results — phrase too long
search_directory("RWA")                              # → seed slugs
search_directory("tokenization")                     # → more seeds
search_directory("stablecoin")                       # → adjacent seeds

# Pass 3 — grow from the strongest seed
find_similar_projects_semantic(slug="<rwa-seed-slug>", limit=10)

# Pass 4 — narrative + sources (also rescues the long phrase from Pass 2)
search_content_semantic(query="real world asset tokenization on Stellar",
                        types=["articles"], limit=8)   # → e.g. "tokenization-basics"
get_related_projects(content_id=<that-article-id>, content_type="article")

# Pass 5 — detail each keeper
get_project(slug="<slug>", compact=true)

From there, assemble the landscape: group the projects by the five categories, note their tags and regions, list the cited content under "context", and call out which categories are crowded vs empty (the gaps).

Output: the landscape

Produce a landscape grouped by category, with a project table (slug + tags + region + website), a cited-content section (one-line takeaway + source url per item), and an explicit gaps / white-space note. Use the full template at reference/landscape-template.md.

A skeleton:

• Sector: <topic> — one-sentence scope.
• Per category (Applications / Developer Tooling / Education & Community / Financial Protocols / Infrastructure & Services): a table of projects | project | slug | tags | region | website |.
• Context & sources: bullet list, each takeaway — <url> from Pass 4.
• Gaps / white space: categories or capabilities with thin or no coverage.
• Go deeper: route to ../stellar-project-dossier/SKILL.md for any single project, ../stellar-integration-finder/SKILL.md to pick a dependency, or ../stellar-builder-quickstart/SKILL.md to turn the gap into a build plan.

Gotchas

• search_directory is keyword/ILIKE — SHORT terms ("wallet", "RWA", "oracle") hit; long NL phrases return 0. For concept discovery use search_content_semantic / find_similar_projects_semantic.
• The content tools return AI SUMMARIES + a source url, not full article text — cite the url.
• get_related_projects is content→projects (needs content_id + content_type); it is NOT project→project. Use find_similar_projects_semantic for that.
• get_project default is a large full row; pass compact=true for the slim shape.
• get_tags_vocabulary = content tags; get_project_tags_vocabulary = project tags.
• If a tool stops appearing after a server update, reconnect the connector (see ../lumenloop-mcp-connect/SKILL.md).

Further reading

• LumenLoop MCP overview and full tool catalog → ../lumenloop-mcp-connect/SKILL.md.
• Stellar developer docs (build references) → https://developers.stellar.org.