June 4, 2026 · Reference post · MCP surface
The 217+ DST factors that make or break an MCP server's agentic readiness
Every audit Appkitekt AI runs on an MCP server or tool-calling API rolls up to one score on one frame: Discovery · Selection · Trust. Below is a preview of the full list — only a subset of the 217 factors we score across registry presence, machine-readable manifests, tool schema quality, parameter ergonomics and trust signals. Run an audit to unlock every detail.
Why DST for MCP
A score is only useful if it tells you where you're losing. DST gives every MCP factor a home in one of three places an agent's tool-calling decision can break:
- Discovery (30%) — can the agent find your server and identify it as a candidate tool? Registry listings, /.well-known manifests, tools/list completeness, OpenAPI hygiene, docs findability.
- Selection (45%) — once you're in the tool set, do you get called? Tool name and description craft, JSON Schema rigor, parameter ergonomics, deterministic outputs, latency and token efficiency.
- Trust (25%) — will the agent invoke you on the user's behalf? Auth simplicity, scope minimization, rate limit transparency, security posture, predictable failure modes.
MCP · 217 factors
MCP surface — DST factor preview
What agents read on MCP servers, tool catalogs, OpenAPI specs, /.well-known manifests and partner registries across Claude, ChatGPT, Cursor, Copilot and beyond.
Discovery · weight 30% · 47 factors
Can an agent find this MCP server / API and recognize it as a candidate tool?
Identity & Provenance · 10
- Stable canonical URL for the MCP server / API base (no churn across versions)
- Server name + vendor name match the legal brand and trademark owner
- Domain ownership matches the marketing site and support URLs
- Legal entity registered in a transparent jurisdiction
+ 6 factors hidden
Registry & Catalog Presence · 15
- Listed in the official Model Context Protocol server registry
- Listed in Anthropic / Claude tool catalogs and partner directories
- Listed in ChatGPT, Cursor, Windsurf, Continue, Cline and other MCP-client catalogs
- Listed in Smithery, mcp.so, Glama, PulseMCP and other community indexes
- Listed in Postman, RapidAPI, APIs.guru, public API directories
- Listed on OpenRouter / function-calling tool marketplaces where applicable
+ 9 factors hidden
Machine-Readable Manifests · 12
- MCP server exposes accurate server metadata (name, version, capabilities)
- tools/list returns a complete, current catalog of tools the agent can call
- resources/list and prompts/list populated where applicable
- OpenAPI 3.x specification published and reachable without auth
- OpenAPI spec passes validation (Spectral / Redocly rules clean)
+ 7 factors hidden
Documentation Findability · 10
- Quick-start that calls a real tool in under 60 seconds
- Server-side rendered docs (full content visible to non-JS crawlers)
- Markdown-readable surface for AI fetchers (.md route variants or clean semantic HTML)
- Versioned docs with clear current vs. legacy split
+ 6 factors hidden
28 factors in this section are only visible in a full ARO audit.
Run audit to unlockSelection · weight 45% · 101 factors
When an agent compares N tools for the same task, why does it call yours?
Tool Schema Quality · 12
- Tool name is a clear verb-object phrase (e.g. search_orders, not v2Op)
- Tool description ≤ 1024 chars yet covers when to use vs. when not to use
- Description includes concrete intent examples the model can pattern-match
- JSON Schema for parameters: every field typed, no any / oneOf-of-string hacks
- Required vs. optional parameters explicit and minimal
+ 7 factors hidden
Parameter Ergonomics · 10
- Parameter names are self-explanatory English, not internal acronyms
- ISO-standard formats (ISO-8601 dates, RFC-3339 timestamps, ISO-4217 currency, ISO-639 language)
- IDs accept either canonical ID or human-friendly slug where ambiguity is low
- Pagination parameters consistent across tools (limit/offset or cursor — not both)
+ 6 factors hidden
Token & Payload Efficiency · 10
- Median tool response under category p50 token count
- Response includes only the fields the agent asked for (field projection)
- Verbose vs. minimal response modes available
- No HTML / boilerplate in JSON responses
+ 6 factors hidden
Performance & Reliability SLOs · 10
- Time-to-first-byte p50 < 200 ms, p95 < 1 s on read tools
- End-to-end p95 latency competitive with category median
- Cold-start latency disclosed and bounded
- Throughput / RPS limits documented per plan
+ 6 factors hidden
Capability Coverage & Completeness · 10
- Tool catalog covers the full job-to-be-done, not just read paths
- Write tools cover every reversible user action
- Search / list tools expose the same filters available in the UI
- Webhook / event subscription for state changes
+ 6 factors hidden
Per-Agent Client Compatibility · 11
- Claude Desktop: drop-in claude_desktop_config.json snippet published
- ChatGPT: MCP server enabled in ChatGPT's Developer Mode / connector catalog
- Cursor: mcp.json snippet published, tools render in the Cursor MCP panel
- Windsurf / Codeium: server callable from the Cascade tool surface
- Continue: config block published for the Continue VS Code / JetBrains extension
+ 6 factors hidden
Auth & Onboarding Friction · 10
- OAuth 2.1 with PKCE supported (no legacy implicit flow)
- Fine-grained, revocable scopes (no all-or-nothing tokens)
- Bearer-token auth path for headless agent runners
- API keys scopeable per tool / per resource, not global
+ 6 factors hidden
Error Semantics · 10
- Error responses include machine-readable code + human reason + remediation
- RFC 7807 problem+json or equivalent typed error envelope
- Validation errors point at the offending parameter path
- Permission errors name the missing scope, not just 'forbidden'
+ 6 factors hidden
Cost & Pricing Transparency · 9
- Per-call / per-token cost stated on the public pricing page
- Pricing parseable as structured data (table / JSON-LD), not images
- Cost preview returned in response headers (X-Usage-Cost) where feasible
- No surprise fees for retries, webhooks, or egress
+ 5 factors hidden
Selection Signals & Social Proof · 9
- Brand preference in agent benchmark prompts for the category
- Selection rate in public agent leaderboards (Tau-Bench, SWE-Bench, etc.)
- Mindshare in dev forum recommendations (Reddit, HN, Discord)
- Case studies of agents successfully running the tool at scale
+ 5 factors hidden
59 factors in this section are only visible in a full ARO audit.
Run audit to unlockTrust · weight 25% · 69 factors
Will an agent stake the user's credentials and reputation on this tool?
Authentication & Secret Handling · 10
- TLS 1.2+ enforced; no plaintext fallback
- HSTS preload on the API and docs domains
- OAuth refresh-token rotation enabled
- PKCE required for public clients
+ 6 factors hidden
Permissions & Consent Model · 8
- Scopes named clearly (read:orders, write:invoices) — no opaque level numbers
- Granular per-resource / per-tool scopes, not surface-wide
- Reversible operations bounded by an undo window
- Human-approval checkpoints documented for high-stakes tools (payments, deletes)
+ 4 factors hidden
Security Posture & Certifications · 10
- SOC 2 Type II report available under NDA
- ISO 27001 / 27017 / 27018 certification where applicable
- GDPR + CCPA / CPRA compliance disclosed with DPA template
- HIPAA BAA available for health data tools
+ 6 factors hidden
Data Handling Honesty · 9
- Training-data opt-out for prompts / responses sent through the tool
- No silent retention of payloads beyond stated TTL
- Region pinning honored (EU data stays in EU)
- Subprocessor list published and versioned
+ 5 factors hidden
Behavioral Honesty & Output Quality · 8
- Tool actually does what its description claims (no overpromising)
- Hallucination / spurious-result rate measured and published
- Outputs cite sources where the tool returns retrieved facts
- Deterministic on identical input (or randomness scoped to a declared seed)
+ 4 factors hidden
Versioning & Stability · 8
- Semantic versioning on the API + MCP server
- Public deprecation policy with minimum N-month notice
- Breaking changes gated behind a new major version
- Changelog parseable by humans and by agents (machine-readable feed)
+ 4 factors hidden
Operational Transparency · 8
- Public status page with per-region, per-tool health
- Incident post-mortems published with timeline + remediation
- Mean-time-to-detection / mean-time-to-recovery disclosed
- Maintenance windows announced ≥ 7 days ahead
+ 4 factors hidden
AI-Native Trust (forward-looking) · 8
- Capability manifest enumerating exactly what an agent may invoke
- Cryptographic attestations / signed responses for high-stakes tools
- Provenance chain for any retrieved or generated content
- Multi-agent coordination support (session isolation, scoped memory)
+ 4 factors hidden
37 factors in this section are only visible in a full ARO audit.
Run audit to unlockNext step
Unlock the complete factor list.
Run an ARO audit to see every signal scored for your MCP server.