landscape-mcp

A local-stdio MCP server that exposes the AI Agent Infrastructure Landscape catalog as a structured query interface for MCP clients (Claude Desktop, Claude Code, Cursor, Cline, etc).

Read-only. Wraps data/landscape.json (912 records × 85 columns) plus data/landscape.edges.json (528 typed edges). Submissions still go through the /submit route and GitHub Issues.

Why this exists

The catalog's edge over LLM synthesis is structured, maintained data. This MCP server makes that data directly addressable from any MCP-capable client — search, compare, drill into edges, surface coverage gaps. Instead of competing with LLMs, it becomes a tool LLMs use.

See issue #48 for context.

Quickstart — local clone (recommended today)

git clone https://github.com/MrPeppersDev/agent-infrastructure-landscape.git
cd agent-infrastructure-landscape/mcp
npm install
npm run build
node dist/test-tools.js   # smoke test — calls every tool, prints results
npm start                 # runs the server on stdio (Ctrl-C to exit)

To wire your local clone into Claude Code:

claude mcp add landscape -- node /absolute/path/to/repo/mcp/dist/server.js

Restart Claude Code; the 12 tools below are now available.

To wire your local clone into Claude Desktop, edit your config:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Add:

{
  "mcpServers": {
    "landscape": {
      "command": "node",
      "args": ["/absolute/path/to/repo/mcp/dist/server.js"]
    }
  }
}

Restart Claude Desktop. The landscape server should appear in the MCP menu.

Quickstart — npx

The package is published to npm with a bundled dataset snapshot, so the install collapses to one line:

# Claude Code
claude mcp add landscape -- npx -y [email protected]

// Claude Desktop config
{
  "mcpServers": {
    "landscape": {
      "command": "npx",
      "args": ["-y", "[email protected]"]
    }
  }
}

Versions

npm version tracks the bundled data-vX.Y.Z tag; bump on dataset releases only. Pin to a specific version (e.g. [email protected]) when you care about reproducibility — otherwise npx -y landscape-mcp picks up whichever dataset shipped most recently.

Tools

All 12 tools are read-only. Inputs are JSON; outputs are JSON-serialisable.

search_records(query, section?, tier?, limit?)

Free-text substring search over id, name, desc cell, and claims cell. Case-insensitive. Returns compact summaries — use get_record for full details.

• query — required search string
• section — exact match on primary section name (use list_sections for valid values)
• tier — 1 (battle-tested) through 5 (theoretical / informal)
• limit — default 25, max 200

get_record(id)

Full record by stable id (all 68 cells, taxonomy, section memberships). Returns null if the id isn't in the catalog. Find ids via search_records.

find_related(id, edge_type?)

Every edge touching the given record id, in both directions, with summaries of the records on the other end. Optionally filter by edge type:

• built-on — built on top of another product
• runtime-dependency — declared dependency at run time (e.g. on a foundation model)
• extends — extends another product's capability
• forks — code fork
• integrates-with — supports integration with the target
• competes-with — competes for the same customer
• inspired-by — design / research influence
• cites — academic citation (from Semantic Scholar; surfaces isInfluential flag)
• same-team-as — same originating team / org
• succeeds — successor product / paper

coverage_summary(dimension)

Catalog-wide per-feature counts for one of four dimensions:

• observability — 7 boolean obs-* columns (LangSmith, OpenTelemetry, Datadog, Helicone, Weave, Langfuse, Arize)
• cost — 7 cost-* columns (token budget, prompt caching, semantic caching, batch API, model routing, streaming-only, cost attribution)
• eval — 7 eval-* columns (LangSmith Evals, Braintrust, W&B Agent Eval, Helicone Evals, custom test harness, human-loop, prod traffic replay)
• benchmark — per-section counts for benchmark-bearing sections

Percentages are denominated against the analyzed set (records with ≥1 yes/no signal), not the full catalog — so a low overall backfill rate doesn't artificially deflate every feature.

compare(id_a, id_b)

Side-by-side comparison of two records across every column. Returns each cell pair plus a match flag for trivial diffability. Also includes any direct edges between the two records.

list_sections()

Alphabetised list of unique primary section names. Use as the section filter on search_records.

recent_changes(since?)

Records whose latest-release (or created, if no release) is on or after the given date. Date is the shipping date of the upstream product, not when the catalog was updated. Accepts YYYY, YYYY-MM-DD, or full ISO timestamps. Omit since for the 20 most-recent.

find_eval_orphans()

Products with observability ≥1 tool but ZERO eval tools, where the absence has been verified (≥1 eval cell carries a yes/no/researched signal). Surfaces the structural failure mode from LangChain's State of Agent Engineering 2025 survey: 89% of practitioners have observability adopted, only 52% have evals — a 37-point gap that costs builders real production trust.

find_substrate_risk(substrate)

Records that runtime-depend on a given substrate. Useful for blast-radius questions ("how many products break if OpenAI changes pricing?") and lock-in analysis ("which framework is most coupled to a single LLM provider?").

Substrate resolves by exact id first, then by case-insensitive name/id substring (picks the match with the most incoming runtime-dependency edges). Examples:

• Anthropic → resolves to Anthropic Claude foundation models, ~62 dependents
• OpenAI → resolves to OpenAI GPT family, ~49 dependents
• MCP → resolves to the MCP spec, ~34 dependents

find_by_decay_cause(cause)

Records whose decay_cause field matches the given cause. Surfaces the mortality forensics from T3-1 — useful for "which products in this space were acquired vs ran out of funding vs were archived on GitHub?"

Enum values: acquired, pivoted, unfunded, lost-benchmark-race, superseded, archived, unknown.

predict_citations(record_id)

Wang-Song-Barabási log-normal citation breakout prediction for a single academic-paper row. Returns the fitted parameters (λ immediacy, μ peak time, σ longevity), predicted asymptote with 90% CI, 10-year prediction, R², phase (pre-growth / growth / saturation / underfit-data), and breakout probability. Methodology: Wang, Song & Barabási, Science 342:127 (2013).

Returns an error if the record is not an academic paper or its citation trajectory is missing. Use list_breakouts for the leaderboard of top predicted breakouts.

list_breakouts(limit?)

Top-N predicted citation breakouts — growth-phase papers with the largest asymptote/observed ratio (still-growing fastest), restricted to rows with real bucketed citation trajectories. Use as a watchlist for which papers to track for citation impact over the next 12-24 months.

• limit — default 15, max 50

Data freshness

The server loads data/landscape.json and data/landscape.edges.json from the repo at process start and caches in memory. Restart the server (or your MCP client) to pick up dataset updates. The data files are semver-tagged via GitHub Releases (data-v1.0.0 onwards) — see data/README.md.

Implementation notes

• Pure query layer. src/tools.ts contains every query as a pure function: (records, edges, args) → result. The MCP server (src/server.ts) is a thin wrapper that registers each function with the SDK. The CLI in issue #49 imports the same tools.ts directly.
• No mutations. Submissions go through the /submit web form → GitHub Issues. This is intentional: an editorial review step is what makes the catalog trustworthy, and a write tool would bypass that.
• No network. All queries run against the local JSON snapshot. No external API calls.

Constraints

• Local stdio only. No remote / hosted MCP, no SSE, no streaming.
• Read-only. No write tools.
• No auth. Not needed for local stdio.

Related work

• #35 — standalone versioned dataset (T0-1)
• #37 — claim-tier provenance (T0-3)
• #46 — runtime-dependency edges (T2-1) — powers find_substrate_risk
• #49 — local CLI tool (consumes tools.ts directly)