MrPeppersDev/agent-infrastructure-landscape

An open, comparative catalog of the tools developers use to build, deploy,
and operate **autonomous AI agents** — the software systems (Claude Code,
Cursor, AutoGen, LangGraph, Mem0, Zep, etc.) that take instructions, plan,
call tools, remember things across sessions, and act in the world.
Each entry is tracked across **85 attributes** (license, maturity tier,
deployment model, MCP / A2A protocol support, observability stack,
compliance posture, latency, pricing, and dozens more), with **typed
relationships** between systems (built-on, extends, competes-with,
cites…). Today the catalog covers **912 systems and 528 relationships**.
The goal is straightforward: give builders, researchers, and analysts a
single place to compare what exists, see what's actually shipping vs.
just published, and spot the gaps.
Start here
- **Browse the catalog →**
— sortable, filterable table with full per-system detail pages, side-by-
side comparisons, lineages, leaderboards, and "best of" lists (e.g.
open-source agent memory,
MCP-enabled systems).
- **Headline findings →** five short, citable stories distilled from the
catalog. Read them as cards on the site
or in source: docs/FINDINGS.md.
- **Raw data →** data/landscape.json +
data/landscape.edges.json. Licensed
**CC-BY-4.0** — use it however you like, just credit the source.
What's in the data
Each record is a JSON object with a stable ID, display name, tier
(1 battle-tested → 5 theoretical), one or more section memberships
(e.g. "Dedicated memory layers", "Framework-embedded memory",
"Vector-database infrastructure" — 34 sections in total), and the 85
attribute cells. Every cell carries:
- The claim itself (free-text value).
- A source URL (where the claim came from).
- A status (real-data, estimate, not-applicable, no-data, etc.).
- A **provenance tier** (T1 auto-verifiable from a GitHub URL, T2
resolvable source URL required, T3 estimate).
- A last-verified date for high-volatility cells.
Sources include curated lists (Agent-Memory-Paper-List,
Awesome-GraphMemory), survey papers, benchmark leaderboards (LongMemEval,
LoCoMo, ConvoMem), vendor sites, academic venue pages, and targeted
research-agent sweeps. Claims are vendor-stated unless otherwise marked.
Honest coverage confidence is roughly **88–92%** — known gaps are tracked
in PLAN.md. Full schema: docs/SCHEMA.md.
Use it in your own tools
**MCP server** — query the catalog from any Model Context Protocol
client (Claude Code, Claude Desktop, Continue, etc.). Nine read-only
tools: search, get-by-id, edge traversal, coverage stats, side-by-side
comparison, recent changes, eval-orphan detection, substrate
blast-radius analysis. See mcp/ and
mcp/README.md.
**CLI** — the same nine queries as a landscape tool.
Text output by default, --json / --csv for machine-readable use.
See cli/ and cli/README.md.
**Web app** — SvelteKit static export in web/, deployed to
GitHub Pages on every push to main.
How the catalog stays honest
Three things keep the data trustworthy:
1. **Per-cell provenance.** Every claim points at the source URL it was
sourced from, with a tier marking how verifiable the claim is. You can
audit any entry by walking the citation.
2. **Automated freshness checks.** A weekly job
(.github/workflows/staleness.yml) flags rows whose upstream repo
has gone quiet beyond the freshness SLA defined in