Capability Radar #001 — MCP discovery is the bottleneck

Status: sample public issue / MVP validation artifact
Audience: AI agent builders, automation operators, MCP users, and technical founders
Core question: which agent capabilities are useful enough to test now?

Top recommendation

Test: Glama MCP Registry / Inspector / Gateway

Category: MCP registry, inspector, hosted gateway
Verdict: Test
Why it matters: Glama is attacking the capability discovery problem directly: browse MCP servers, inspect tools in-browser, and put a gateway in front of multiple servers with logging and access control. For builders, the useful bit is not only directory search; it is the try-before-install workflow and gateway concept.
Builder use case: evaluate candidate MCP servers before adding them to an agent runtime; compare tool descriptions and schemas; test whether a server exposes the capability you actually need.
Friction: external hosted platform; production use requires trust in gateway/security model.
Recommendation: test with non-sensitive MCP servers first. Do not route private credentials or internal systems until the security model is reviewed.

Source: Glama

Test this week

1. Official MCP Registry

Category: MCP server registry
Verdict: Test
Why it matters: The official registry is the closest thing to a canonical MCP discovery surface. The modelcontextprotocol/servers GitHub repo now explicitly points users to the registry for published servers and clarifies that the repo is mainly for reference implementations.
Builder use case: use it as the default first stop for MCP server discovery before searching random GitHub repos.
Friction: registry coverage may lag the wider ecosystem; still verify individual server maintenance and security.
Recommendation: make it the baseline source in the weekly monitoring pipeline.

Sources: MCP servers repo, MCP Registry

2. Agent Skills / SKILL.md standard

Category: Agent skill standard
Verdict: Test
Why it matters: Agent Skills package repeatable workflows as folders with SKILL.md, scripts, templates, and references. The pattern is spreading across multiple agent clients. This makes capabilities more portable than one-off prompt snippets.
Builder use case: convert repeated workflows — code review, report generation, MCP setup, QA procedures — into reusable skills that load only when relevant.
Friction: ecosystem discovery is immature; quality varies; skills can encode unsafe workflows if not reviewed.
Recommendation: start by writing one internal skill for a painful repeated workflow, then evaluate public skill repos later.

Sources: Agent Skills overview, Agent Skills specification

3. Smithery MCP directory

Category: MCP server directory / publishing surface
Verdict: Test
Why it matters: Smithery is one of the visible discovery and publishing surfaces for MCP servers. It is useful for finding servers by category and seeing how MCP publishers present themselves.
Builder use case: discover candidate MCP servers and study listing metadata before deciding what to install or replicate.
Friction: directory presence is not proof of maintenance, safety, or production quality.
Recommendation: use for discovery, not trust. Pair every candidate with GitHub activity, docs quality, and local sandbox checks.

Source: Smithery MCPs

Skip for now

1. Random MCP servers with weak docs

Category: MCP server long tail
Verdict: Skip
Why it matters: The ecosystem is large and duplicative. An MCP server with no clear install path, unclear credential handling, or stale commits should not be added to a real workflow just because the capability sounds useful.
Recommendation: require minimum trust gates: recent maintenance, clear README, explicit auth model, local test path, and small blast radius.

2. Production routing through unreviewed hosted gateways

Category: MCP hosting / gateway
Verdict: Skip for production; test in sandbox
Why it matters: Gateways are powerful because they centralize access, logging, and credentials. That also makes them sensitive infrastructure.
Recommendation: sandbox first; never route private company systems or paid API keys without a review.

3. “Agent marketplace” claims without real buyer workflows

Category: Agent marketplaces
Verdict: Skip
Why it matters: Many marketplace concepts are narrative-first. The useful near-term problem is not agents shopping autonomously; it is human builders discovering maintained capabilities with clean install and trust signals.
Recommendation: watch marketplaces that solve discovery, verification, install, and billing friction — not just listing pages.

Watch closely

1. x402 and agent-native paid APIs

Category: Payment/access protocol
Verdict: Watch
Why it matters: x402 uses HTTP 402 Payment Required style flows to let agents pay for resources/API calls. It may become a useful primitive for agent-accessible paid capabilities.
Builder use case: pay-per-call APIs, datasets, and microservices without traditional account signup.
Risk: still early; wallet custody, spend limits, accounting, fraud, and compliance need careful handling.
Recommendation: track ecosystem explorers and examples; do not put autonomous spend in production yet.

Sources: x402scan, Chainalysis x402 analysis

2. Google AP2 / agent payment authorization

Category: Agent payments / authorization protocol
Verdict: Watch
Why it matters: AP2 focuses on secure authorization and mandates for agent-led payments across payment methods. That is directly relevant to agent capability marketplaces if agents eventually purchase tools, data, or services.
Recommendation: watch for developer-friendly implementations and MCP/A2A integration examples.

Source: Google Cloud AP2 announcement

3. Capability metadata quality

Category: Discovery infrastructure
Verdict: Watch
Why it matters: Tool descriptions, schemas, installation metadata, trust signals, and usage examples determine whether agents and humans can select the right capability. Discovery quality may become the real competitive surface.
Recommendation: track registries that expose machine-readable metadata, quality scores, examples, and security notes.

Builder experiment

Run this 30-minute test:

  1. Pick one repeated workflow your agent performs.
  2. Search the official MCP registry, Smithery, and Glama for one relevant capability.
  3. Reject anything without clear docs/auth/security notes.
  4. Install only in a sandbox profile/project.
  5. Record:
    • install time
    • credentials required
    • tool names/descriptions
    • whether the agent selected the right tool
    • whether the output saved time

If the capability does not save time in one hour, remove it.

Radar queue for issue #002

  • MCP server security scoring
  • Best sources for Agent Skill discovery
  • x402 paid API examples
  • Agent marketplace reality check
  • Capability metadata schema proposal

CTA

Suggest one capability to review next: MCP server, Agent Skill, registry, payment/access protocol, or marketplace.

Interim contact: leodavinciad@users.noreply.github.com with subject Capability Radar suggestion.