AI symbol and footprint generators in 2026: who publishes accuracy?

You would expect a tool that turns a datasheet into a footprint to tell you how often it gets the footprint right. A wrong pad pitch costs a board spin; a mirrored pinout costs two. Accuracy is the product.

As of October 14, 2026, not one of the eleven commercial tools in this post publishes an accuracy number. The only measured figure we could find for AI-generated symbols and footprints comes from an academic paper: 86 percent for symbols and 80 percent for footprints on a dataset of 1,000 components (SFgen, arXiv 2607.19767).

That is a useful number, and an uncomfortable one. One footprint in five wrong, on the best measured pipeline, is not a tool you let commit to a library unattended. It is a tool that writes a first draft. This post compares what each generator actually does, what it costs where that is public, whether your assistant can call it, and then gives the three-step routine that makes "first draft" safe.

The comparison

"Accuracy published" means a numeric claim about symbol, footprint or pinout correctness on the vendor's own site or in a linked publication. Pricing is listed only where we found it on a page we could read. "API or MCP" is what the vendor advertises, not what community wrappers exist.

Tool	What it generates	Pricing (where public)	API or MCP	Accuracy published
Stillwind	Repairs footprints against the datasheet (IPC-7351B, KLC); free Stillwind Search	On request	No	No
PCBPartz	Symbol, footprint and 3D from a datasheet (AI Symbol Expert); 13M+ IPC-7351 parts	Not found	Not advertised	No
Librizer	Symbol, footprint and device map for Fusion and KiCad from a PDF	$0.50 to $2.00 per part, $3 free credits	Not advertised	No
Flux	Browser eCAD with a Copilot agent: plan and execute design, auto-layout, sourcing-aware with live pricing	Not in our sources	Not advertised	No
FootprintAI	Terminal agent: libraries and schematics from plain English (Eagle, KiCad)	Waitlist, "launching soon"	Not advertised	No
CELUS	Schematic and BOM from text requirements for Altium, Cadence, Siemens, Fusion, KiCad	Not public	Agents consume manufacturer data over MCP	No
Circuit Mind	Schematic, BOM, FMEA and derating reports from requirements	Plans, prices not public	Not advertised	No (peer-reviewed LANL case study, not an accuracy figure)
PartGenie	Part search, datasheet Q&A, alternatives, BOM analyzer, KiCad 8 symbol, footprint and 3D	Not found	MCP listed as a channel	No
Zenode	AI component search over "40M+ part sources", datasheet Q&A	Not found	Not advertised	No
SnapMagic Copilot	Circuit auto-complete, natural-language schematic edits, BOM optimization with live stock	Not on the page we checked	Not advertised	No
datasheets.md	Pin tables, footprints, KiCad symbols, datasheet extraction with provenance	Free; Builder $10/user/mo; Startup $250/team/mo; Enterprise	Official MCP server, REST API, OAuth 2.0	No
SFgen (paper)	Symbols and footprints via an agentic multimodal-LLM flow	Research	n/a	Yes: 86% symbol, 80% footprint, 1,000 components

Two reference libraries most engineers re-check against, neither of which claims to use AI:

Library	What it is	Pricing	Source
Ultra Librarian	16M parts, symbols, footprints and 3D in 30+ CAD formats; part of Cadence	Free to register	ultralibrarian.com
SamacSys / Component Search Engine	Free ECAD model library; SamacSys joined Supplyframe, which Siemens acquired in 2021. The site blocked our fetch, so current product status is unverified	Free	samacsys.com, Siemens

The generators, one by one

Stillwind positions itself as "librarian as a service." Rather than generating from scratch, it repairs existing footprints against the datasheet and checks them against IPC-7351B and the KiCad Library Conventions, with two AI systems cross-checking every value. There is a free search product; pricing for the repair service is on request, and no API or MCP is advertised (stillwind.ai). The cross-check design is the closest any vendor comes to addressing accuracy structurally, but no number is published.

PCBPartz claims more than 13 million IPC-7351 components and launched AI Symbol Expert on February 24, 2026, which turns a datasheet into a symbol, footprint and 3D model (Electronics For You, pcbpartz.com). No numeric accuracy claim.

Librizer, from Leacron Inc. in Tokyo, takes a datasheet PDF and produces a symbol, footprint and device map for Fusion and KiCad. It is the one tool here with fully public per-part pricing: $0.50 to $2.00 per part with $3 in free credits (librizer.app). A hands-on write-up by an independent reviewer concluded the output is a first draft that still needs review (dev.to), which matches our advice below.

Flux (Defy Gravity Inc.) is a browser eCAD tool rather than a library generator, with a Copilot agent that plans and executes design steps, auto-routes, and is sourcing-aware with live pricing and stock. It raised $37 million in a Series B led by 8VC, announced February 27, 2026 (SiliconANGLE). Its BOM export is one of the inputs Next Everything reads, see below.

FootprintAI is a terminal agent that generates libraries and schematics from plain English for Eagle and KiCad. As of our check it is a waitlist ("launching soon"), so there is nothing to measure yet (footprintai.net).

CELUS, from Munich, runs a multi-agent pipeline (orchestrator, planner, checker, verifier) from text requirements to a schematic and BOM for Altium, Cadence, Siemens, Fusion and KiCad. It is notable here for one reason: manufacturers' product data is exposed to its agents over MCP, which makes it one of the few design tools in this list that speaks the protocol at all (celus.io). Its Series A was in July 2022.

Circuit Mind, from London (founded 2018), generates a schematic, BOM and FMEA or derating reports from requirements. It has a peer-reviewed case study with Los Alamos National Laboratory, which is evidence of real use, not an accuracy figure. Plans exist but prices are not public (circuitmind.io).

PartGenie is an AI hardware assistant: part search, datasheet Q&A, an alternative finder, a BOM analyzer and a KiCad 8 symbol, footprint and 3D generator. MCP is listed as a channel. One thing to note when you read its site: the home page says "56M+ parts" and the FAQ says 35 million (partgenie.ai).

Zenode is AI component search over "40M+ part sources" with datasheet Q&A, shown on Hacker News on September 22, 2025. The discussion there is a good snapshot of what engineers worry about with this category: whether to trust AI summaries of a datasheet, response times around a minute, and broken links (Hacker News).

SnapMagic Copilot sits on top of the SnapMagic (formerly SnapEDA) model database. It auto-completes circuits, takes natural-language schematic edits, optimizes a BOM against live stock from Mouser and Digi-Key and offers one-click BOM purchase, with integrations for Altium, Fusion, KiCad, Zuken and Cadence (snapmagic.com). The company rebranded in October 2023 with the Copilot launch. We did not find pricing on the page and no API or MCP is advertised.

datasheets.md (Plinsburg Tech, Poland) is less a generator than a data layer: pin tables, footprints, KiCad symbols and datasheet extraction where every value carries provenance back to the PDF page, table and cell. It has an official MCP server, a REST API and OAuth 2.0. Pricing is public: Free (rate-limited), Builder at $10 per user per month, Startup or Partner at $250 per team per month, Enterprise custom, with overage at $0.002 per call (datasheets.md, pricing). Provenance is the feature to look for when you cannot get an accuracy number: it tells you where to look to check.

The one measured number

The SFgen paper, "Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation" (arXiv 2607.19767, July 2026), describes a multimodal-LLM agentic flow that reads datasheets and generates symbols and footprints. On a dataset of 1,000 components it reports 86 percent symbol accuracy and 80 percent footprint accuracy (arXiv).

Three cautions before you quote it. It is a research pipeline, not any product in the table, so it says nothing about any vendor's number, better or worse. The dataset is 1,000 components, and the paper's definition of "accurate" is its own. And the gap between the two figures, symbols at 86 and footprints at 80, matches what engineers report anecdotally: pin names are easier to read off a PDF than pad geometry is.

If a vendor in the table publishes a number on a comparable dataset, we will add it here. Until then, the honest statement is: the best measured pipeline gets one footprint in five wrong.

How to use a generated footprint safely

Treat every generated library part as a draft from a fast, slightly careless colleague. The routine is three steps, and the third is the one people skip.

1. Verify against the datasheet, not against the tool. Open the mechanical drawing and check, in this order: pin count and numbering direction, pad pitch, pad size, courtyard, and polarity or pin 1 marking. For symbols, check the pin names against the pin table and the power pins against the recommended operating conditions. A provenance link, where the tool offers one, makes this a two-minute job. Where it does not, print the drawing.

2. Compare with a reference library. Ultra Librarian and Component Search Engine are the libraries most teams already re-check against, and a diff between a generated footprint and a manufacturer-supplied one catches most pitch and courtyard mistakes. Where the two disagree, the datasheet decides.

3. Check that you can buy the part before you commit to the footprint. This is the step the generators do not cover. A perfect footprint for a part with no stock, a 40-week lead time or a "not recommended for new designs" flag is a board spin waiting to happen. Design with what you can buy makes the case in full.

That third step is where Next Everything fits. It is an MCP server that your assistant (Claude, Claude Code, Cursor, Codex, VS Code and others) connects to with a personal key; nx_search_parts returns live stock and prices by part number, parametric spec or plain English, with alternates ranked by compatibility and filters for package, temperature range, RoHS and minimum stock. So the check can happen in the same chat where the footprint was generated: "is this part in stock at 50 pieces, and what are the drop-in alternates in the same package?" nx_scan_project reads KiCad 7 to 9, atopile, tscircuit, Flux BOM exports and Altium or EasyEDA BOM CSVs, resolves every line to an orderable part and flags the ambiguous ones, which is a useful last pass before a library change goes into a release. If you want to see what MCP is first, MCP in five minutes covers it.

Next Everything does not generate symbols or footprints. It also does not publish an accuracy number for part resolution, so we are in the same position as the vendors above on that score. What it does publish is the rule that nothing gets swapped without your OK, that out-of-stock parts come back with a clearly marked replacement, and that your assistant cannot spend money: nx_checkout makes a payment link you open yourself. The service is free: you pay only for the parts in an order you approve.

Summary

• Eleven commercial tools generate or repair symbols, footprints, schematics or BOMs with AI. None publishes an accuracy figure as of October 2026.
• The only measured number is academic: 86 percent symbol, 80 percent footprint on 1,000 components.
• Public pricing exists for Librizer and datasheets.md; the rest is on request, not listed, or not yet launched.
• Official MCP access exists at datasheets.md (component data), CELUS (its own agents) and, as a listed channel, PartGenie.
• Generated footprint, then datasheet, then reference library, then stock check. In that order.

For the wider picture of who sells parts to agents and who does not, What happened to on-demand manufacturing marketplaces has the full list. To put the stock check inside your assistant, sign up, paste one snippet, and ask whether you can buy the part before you keep the footprint. The docs have the setup for each client.