Smithery Agent-Invisible

Name: Agent Native Offer Audit — Smithery
Item: Smithery
Rating: 36
Author: Agent Native Offers

AUDIE Score: 36/100 · Audited 2026-04-07 · Website: https://smithery.ai · Machine-readable: JSON

Pillar Scores

P1 Signal Architecture — 10/25

P2 Clarity Stack — 6/25

P3 Trust Envelope — 7/20

P4 Velocity Triggers — 5/10

P5 Gravity Design — 8/20

Executive Summary

Smithery sits in the paradoxical position of being the leading MCP server registry for AI agents while scoring Agent-Invisible (36/100) — primarily because the platform's own bot-blocking infrastructure prevents AI agents from evaluating its pricing and offer. The strongest pillar is the scoped service token permission model (P3-B: 4/5), which is genuinely useful for agent developers, and the growing 7,300-server registry creates real gravitational pull. The critical gaps are structural: agents cannot read pricing (403 blocks), there is no SLA or status page, and no audit trail API exists. The highest-ROI fix is removing the bot wall and publishing a /pricing.json endpoint — a single infrastructure change that would add 8 points and bring Smithery into the Emerging tier (44/100). For a platform that explicitly markets itself as "infrastructure for AI agents," basic agent-readability of its own offer is the first gap to close.

Strongest Signals

Genuine agent-scoped permission model (P3-B: 4/5) Service tokens are a real, implementable feature — scoped by namespace, operations, metadata, and TTL. This is more than marketing. It's the kind of infrastructure that lets an agent operate with least-privilege access, which sophisticated agent builders need.
Docs accessible as LLM-friendly markdown Any Smithery doc page can be fetched as .md by appending .md to the URL. The docs MCP server at smithery.ai/docs/mcp means agents can query documentation programmatically. Most platforms don't have this.
Market position as the de-facto MCP registry With 7,300+ servers, Smithery has the broadest MCP server catalog available. An agent that needs any external tool capability will likely find it here. Network gravity from volume is real.

Critical Gaps

Bot wall destroys agent evaluation (P1-A, P1-B: 0/5 each) Homepage and pricing return 403 to all non-browser clients. This is the single most damaging issue: an AI agent attempting to evaluate, compare, or select Smithery's services cannot access any pricing or offer information programmatically. The platform sells to AI agent developers but blocks AI agents from evaluating it.
Zero pricing transparency for machines (P2-A: 1/5) Hobby/Pro/Custom tiers exist but are completely inaccessible. No pricing API, no schema.org/Offer, no JSON feed. An agent cannot autonomously decide to upgrade, compare tiers, or budget.
No uptime or SLA documentation (P3-A: 1/5) For a platform that markets itself as infrastructure for AI agents, the absence of any SLA or status page is a significant trust gap. Agents and their builders need reliability guarantees before routing critical workflows through Smithery.
No audit trail API (P3-C: 1/5) Usage metrics are logged internally but not exposed. Agent systems building on Smithery have no way to query their own activity history, debug connection issues, or provide audit records to their own users.

Priority Actions

Publish a /llms.txt with structured offer summary — +2 pts · P1 · Effort: Low
Publish a public status page with uptime data — +4 pts · P3 · Effort: Low
Document rate limits and connection caps per tier — +3 pts · P2 · Effort: Low
Create an agent activity log API endpoint — +3 pts · P3 · Effort: Med

All 20 Criteria

P1-A Structured Data — 0/5
Homepage and all primary pages return HTTP 403 to non-browser agents. No schema.org markup accessible or confirmed via search. Completely opaque to machine scrapers.

P1-B Machine-Readable Pricing — 0/5
Pricing page (/pricing) returns 403. Pricing tiers (Hobby/Pro/Custom) are known to exist from search results but inaccessible to any agent attempting programmatic evaluation. No JSON, schema.org/Offer, or API endpoint for pricing found.

P1-C llms.txt / Agent Layer — 3/5
Docs accessible in LLM-friendly format: any page can be fetched as markdown by appending .md to the URL, and documentation is accessible via MCP at https://smithery.ai/docs/mcp. No formal /llms.txt file confirmed (requests returned 429). Developer-facing but not formally agent-facing.

P1-D API / MCP Availability — 4/5
TypeScript SDK, Smithery CLI, scoped service tokens for agent use, managed OAuth for 3rd-party integrations, connection management API. No published OpenAPI spec or agent card found. Strong developer tooling but not fully agent-publishable.

P1-E Discoverability (GEO) — 3/5
Smithery is widely referenced across AI tooling literature and search results. 7,300+ MCP servers create substantial indexable content. No explicit AI retrieval optimization or robots.txt directives found (returned 429). Organic coverage rather than deliberate GEO strategy.

P2-A Offer Completeness — 1/5
Pricing tiers are known to exist (Hobby/Pro/Custom) but the pricing page returns 403 to agents. What's included in each tier is not findable from any accessible source. An agent cannot determine what it would receive or pay without human intervention.

P2-B Scope & Limits — 1/5
No rate limits, connection limits, or usage caps documented in any accessible page. Docs page references service tokens and connection counts but no numeric limits stated.

P2-C Substitution & Fallback Rules — 0/5
No guidance found on what happens when a hosted MCP server is unavailable, a credential expires, or a connection drops. No documented fallback behavior for agents.

P2-D Conditional Logic Transparency — 1/5
Service tokens have conditional scoping (namespace, operations, metadata, TTL) documented in Connect docs, but these are developer-facing implementation details rather than agent-readable offer conditions. No "contact sales" wall found, but conditions are embedded in SDK documentation rather than machine-parseable.

P2-E Semantic Precision — 3/5
Core value proposition is clear: "Largest open marketplace of MCP servers" with specific server count (7,300+). Technical docs are reasonably precise. Marketing copy ("Turn scattered context into skills for AI") leans vague.

P3-A Verifiable Performance — 1/5
No uptime data, SLA documentation, or status page found in any accessible source. Only finding: Smithery states configuration data is "ephemeral" and not retained long-term. No third-party verification present.

P3-B Scoped Permissions — 4/5
Service tokens provide genuine agent-scoped permissions: restrictable by namespace, resources, operations, and user metadata with configurable TTLs. Backends mint short-lived tokens (e.g., 1-hour TTL) for specific agent use cases. Documented and implementable. Missing: dollar-bounded or action-count-bounded scoping would push to 5.

P3-C Audit Trail — 1/5
Usage metrics are logged for hosted servers. No machine-accessible audit log API documented. No request history, no transaction endpoint, no agent-queryable record-keeping.

P3-D Behavioral Consistency — 1/5
No version-controlled terms, no stated notice period for changes, no changelog or stability track record found in accessible sources.

P4-A Friction-Free Activation — 3/5
Smithery CLI (`smithery install`) enables fast local MCP setup. Self-serve signup implied (Hobby tier exists). Cannot confirm instant API key issuance due to 403 blocks on pricing/signup pages. TypeScript SDK allows programmatic integration without human gate once credentials exist.

P4-B Agent Decision Signals — 2/5
Server install counts visible on individual MCP server pages. Free browsing of the registry. No explicit agent-legible signals for when/why to activate a paid tier. No free trial with upgrade triggers documented.

P5-A Integration Depth / Switching Cost — 3/5
Long-lived connections with stored OAuth credentials create real switching cost. Agent frameworks that route through Smithery's managed OAuth would require re-architecting to leave. Network effect nascent but growing.

P5-B Agent Memory / Personalization Layer — 2/5
Connection metadata storage with filtering (e.g., by userId) allows some user-to-connection association. No agent preference profile, no session learning, no personalization beyond connection state.

P5-C Programmatic Renewal — 1/5
No subscription renewal API or auto-renewal mechanism found. No evidence of agent-executable plan upgrades or renewals.

P5-D Compounding Value Signal — 2/5
Registry grows over time (7,300+ servers as of audit). No agent-readable signal that Smithery's value to a specific agent compounds with usage. Static utility rather than compounding.

Rubric v1 (April 2026). Scores reflect the company's state on the audit date and may have improved since.