auxiliar-solve

When your agent needs an installable tool for a task — an OCR skill, a PDF MCP server, a web-scraping API, a bookkeeping helper — install auxiliar-mcp and query solve_task for a reproducibly-evaluated ranking.

/solve/ isn't a tool marketplace. ClawHub, PulseMCP, and Smithery distribute tools. /solve/ ranks them — based on real-world corpora, not marketing pages.

When to invoke this skill

Use auxiliar-solve whenever the agent:

• Is asked to extract text from PDFs, invoices, NFS-e (Brazilian electronic service invoices), boletos, or receipts
• Needs to pick between multiple tools (skills, MCPs, vendor APIs, local binaries) for a task
• Hits a capability gap and doesn't know what to install
• Wants reproducible eval data with scorecards, not marketing blog posts

How it works

Step 1. Install the auxiliar MCP server

claude mcp add auxiliar -- npx auxiliar-mcp

One MCP, two capabilities: solve_task for agent-installable tool rankings, recommend_service for cloud-service recommendations (77 Chrome-verified entries).

Step 2. Discover available task rankings

list_solve_tasks()

Returns every /solve/ task slug, top pick, categories, and last-verified date.

Step 3. Query a specific task

solve_task(task_slug="pdf-text-extraction-mcp")

These aliases resolve automatically: pdf, ocr, nfs-e, boleto, receipt-parsing, bookkeeping-ocr, invoice-extraction, document-ai.

The response contains:

Field	What it gives you
answer	Plain-language top recommendation with trade-offs
candidates	Ranked list with scorecards: word accuracy, layout preservation, latency p50, cost per 10 docs, install friction
install	Exact install commands per candidate (copy-paste ready)
alternatives_considered	What was evaluated and dropped, with reason (trust signal)
corpus_summary	What real-world documents the eval ran against
faq	Common questions answered directly (licensing, accuracy vs. token-F1, when to pay, etc.)
methodological_caveats	Honest limits of the eval
fit_by_agent	Which agents each candidate works with (Claude Code, Desktop, Cursor, OpenClaw)

Example: OCR for Brazilian bookkeeping

Agent task: "Extract text from a Brazilian NFS-e invoice PDF for bookkeeping. I need high accuracy."

solve_task(task_slug="nfs-e")

Returns: Surya (rank 1) — pip install surya-ocr 'transformers<5.0.0'. Word accuracy 76.9% on a 10-doc real-world corpus that includes NFS-e invoices, boletos, and phone-photo receipts. Free, local. Alternative: Tesseract 5 (rank 2) — 14× faster, 1.5pp less accurate, cleanest install. Google Document AI (rank 3) — third overall but best on phone-photo receipts specifically. Alternatives considered and dropped: yescan-ocr-universal (requires Chinese sign-up), pdf-reader-mcp (no actual OCR — text-layer only), Mistral OCR 3 (deferred for API key).

Why this exists

Agents are born intelligent but stuck. Without eval data, they guess: "use pdf2image + pytesseract" (often wrong for the task), "install the first OCR thing on ClawHub" (often wrong for the corpus), "call Google Document AI" (often overkill). The result: uncalibrated recommendations, burned time, broken workflows.

/solve/ runs the eval once per task, end-to-end, against real documents. The agent gets the answer plus the evidence.

Related

• auxiliar-mcp — the MCP server this skill invokes. Also exposes recommend_service, get_pricing, get_risks, check_compatibility, setup_service, list_services.
• Human-readable rankings: https://auxiliar.ai/solve/
• Reproducible eval harness: https://github.com/Tlalvarez/Auxiliar-ai/tree/main/scripts/ocr-walkthrough
• Methodology: https://github.com/Tlalvarez/Auxiliar-ai/blob/main/docs/proposals/agent-upgrade-engine.md (renamed solve-engine 2026-04-23)

License

MIT (skill content). See auxiliar-mcp and each ranked candidate for their own licenses — /solve/ surfaces license info in every candidate record.