BYOK model router for OCR document analysis, invoice extraction, bank statement extraction, scanned PDFs, and long financial documents.
Server has 10 tools with explicit schemas and descriptions. Naming follows verb_noun pattern (router_route_request, router_list_models, router_execute_request). Schemas are well-structured using Zod with enums for constrained inputs (task_type, provider_filter, tier_filter). However, descriptions are inconsistent in depth, some tools lack actionable context for LLM selection. Parameter descriptions exist but many are generic ('Profile of the document being processed', 'Routing policy constraints') without explaining WHEN to use them or WHAT they control. Output schemas are not documented in the visible code. Error handling and recovery guidance are absent. Tool annotations present but incomplete (readOnlyHint set to false on read-only tools, which is incorrect).
Estimate the cost of a routing decision.
Routes a document OCR/extraction prompt, calls the selected provider when credentials are configured, and returns dry-run output otherwise.
Get learned model performance scores by task type.
Get routing statistics and learning summaries.
Get the history and statistics of a multi-step routing trajectory.
Browse the model catalog with pricing, capabilities, compliance, and filters.
Record the outcome of a routing decision for learning and evaluation.
Tool annotations incorrect: readOnlyHint set to false on read-only tools (router_route_request, router_list_models, router_get_stats, etc.). Should be true for READ_ONLY operations.
Parameter descriptions lack actionable context. 'Profile of the document being processed' and 'Routing policy constraints' do not explain WHEN to provide them or WHAT fields are expected. DocumentProfileSchema and RoutingPolicySchema are complex nested objects but descriptions do not guide LLM on how to construct them.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Call before OCR, PDF parsing, document extraction, validation, or reconciliation. Returns the optimal model, fallback, estimated cost, and reasoning.
Experimental V3 router with Claude-reviewed scoring separation and server-returned applied weights.
V4 OCR/document router with preflight document profile plus benchmark-backed priors from MMR-Bench, MMDocBench, and CC-OCR V2.
Output schemas not documented. Tools return formatted responses (formatDecision, etc.) but no schema is visible in tool registration. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
No error handling or recovery guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. Missing pattern:recovery-guide implementation.
Three routing tool versions (v1, v3, v4) create ambiguity. LLM must reason about which version to call. Descriptions do not clearly differentiate when to use v3 vs v4 (v3 says 'Experimental', v4 mentions 'benchmark-backed priors' but no guidance on which is production-ready).