AI Models for PCB Design: KiCad, Altium, MCP, and Frontier Model Benchmarks
TL;DR
Frontier models can drive KiCad and Altium through MCP servers, computer-use agents, and native AI EDA tools. None of them replace a signed schematic review, netlist check, or DFM gate. For KiCad 10, open MCP stacks such as Konnect and kicad-mcp-pro expose dozens to hundreds of tools; Altium teams are wiring live sessions through eda-agent and related MCP bridges. On published benchmarks, GPT-6 Astra leads BenchCAD and KiCad layout demos; Claude Opus 5.5 leads long-horizon agent indexes; DeepSeek V4.1 Flash wins cost-sensitive terminal and repo repair loops. Pick the model for the job, pick the integration for your EDA, and keep engineering verification in the loop.
Why PCB design moved onto MCP in 2026
General chat can sketch a block diagram. It cannot safely edit your checked-in .kicad_pcb without a controlled bridge. The Model Context Protocol (MCP) gives agents typed tools: run DRC, move a footprint, export Gerbers, query LCSC stock. KiCad forum threads and GitHub repos exploded in 2026 because KiCad 10 shipped a supported IPC API, so Rust and Python servers can edit live projects with undo instead of brittle screen scraping.
Social posts on X and Reddit track the same arc: open-source Altium MCP servers, Konnect for KiCad 10, and caution that vibe-layout still fails ERC in subtle ways. We could not pull live X search in this review; the evidence below comes from vendor posts, JLCPCB case studies, Adafruit, GitHub READMEs, and Techwall’s earlier model notes.
Three ways AI touches your EDA stack
| Pattern | How it connects | KiCad example | Altium example | Best for |
|---|---|---|---|---|
| MCP tool server | Agent calls tools over stdio; EDA stays open | Konnect, kicad-mcp-pro, Seeed kicad-mcp-server | eda-agent, altium-designer-mcp | Incremental edits, DRC, libraries, review |
| Computer-use agent | Model clicks the native UI | GPT-6 Astra KiCad demo; generic OSWorld-style harness | Same, slower, fragile on dialogs | One-off demos, legacy flows without API |
| Native AI inside EDA or cloud | Vendor model or partner API | None built into KiCad 10 | Altium 365 Engineering Assistant; Quilter import | Requirements, placement engines, autorouters |
Flux’s team argues that desktop MCP is like routing pizza through a straw: workable, but high latency per step. For production programs, MCP plus git still beats pure computer use because diffs are inspectable.
KiCad: MCP servers compared
KiCad has no bundled copilot. Your choice is which server, how many tools you expose to the model, and whether writes are opt-in.
| Project | API / transport | Tools (documented) | Write default | Notable limits |
|---|---|---|---|---|
| Konnect (mixelpixx) | KiCad 10 IPC (protobuf/NNG); Rust plugin | 228 tools in 21 toolsets; ~2K token starter vs ~23K if all listed | Write via live IPC | Needs KiCad API enabled; replaces deprecated SWIG path |
| kicad-mcp-pro (oaslananka) | MCP + kicad-cli; DFM/SI helpers | Full review + build profiles | Read-only review profile default; write is explicit | JLCPCB case study picked this for safety |
| Seeed kicad-mcp-server | pcbnew / kicad-cli | Analysis, net trace, codegen | Limited schematic edit (experimental) | Strong for ERC/DRC and pin tracing |
| KiCad AI Assistant (paul356) | In-editor chat + MCP | Schematic/PCB edit set | Write in GUI session | Human watches the canvas |
Adafruit’s Oct 2026 note on Konnect is the clearest public spec on token cost: lazy-loading toolsets matters when Opus or Astra context is expensive. kicad-mcp-pro’s read-only default matters when a wrong pcb_move_footprint call could trash a release branch.
Altium: live session MCP and vendor AI
Altium Designer still owns constraint-driven layout for many enterprises. AI enters two lanes: cloud requirements assistants inside Altium 365, and third-party MCP servers that drive an open Altium session over DelphiScript bridges.
- eda-agent (salitronic): open-source MCP; README claims on the order of 560–600 tools on Altium covering schematic, PCB, libraries, and audits; KiCad and EasyEDA Pro as alternate backends.
- altium-designer-mcp (embedded-society): Rust server focused on library creation with IPC-7351-aware footprint reasoning.
- Altium 365 Engineering Assistant: natural-language requirements in the portal; vendor docs do not label ActiveRoute as generative AI. Layout automation often means exporting to Quilter or DeepPCB, not chatting inside the router.
Reddit threads in early 2026 highlight schematic review through Claude Code plus Altium MCP as the first production use case: ERC/DRC reports in, annotated changes out, human approves.
Frontier models on EDA-shaped benchmarks
No public leaderboard scores “route this 12-layer HDI board.” Proxy benchmarks matter: CAD reconstruction, desktop automation, terminal tasks, and long-horizon agents. Numbers below are from OpenAI’s GPT-6 Astra launch, DeepSeek’s V4.1 Flash card, Artificial Analysis, and Techwall’s prior write-ups. Harnesses differ; treat rows as directional.
BenchCAD and CAD-like reconstruction
BenchCAD (with tools) is the closest published score to “turn views into CAD.” From OpenAI’s September 2026 launch table, also cited in our GPT-6 Astra post:
Source: OpenAI GPT-6 Astra launch. Opus 5.5 is not in this row; use Coding Agent Index for code-heavy layout scripts instead.
Terminal-Bench 4.0 (long-horizon agents)
MCP PCB sessions look like terminal agents: hundreds of tool calls, files, and retries. Terminal-Bench 4.0 separates models that stay coherent over long tasks.
Opus/GPT-6.1 from Artificial Analysis v4.3.2; Astra from OpenAI launch; DeepSeek Oct 2026 API changelog. See our DeepSeek V4.1 Flash post for the full agent table.
AutomationBench (desktop and tool use)
| Model | AutomationBench (published) | Notes for PCB workflows |
|---|---|---|
| Claude Opus 5.5 (AA max) | 69.5% (AutomationBench-AA) | Strong when the agent must drive many UI steps without MCP |
| GPT-6.1 Sol (AA max) | 64.9% | Good balance of cost and tool use in AA harness |
| GPT-6 Astra (OpenAI launch) | 41.4% | Launch table; pairs with Codex harness improvements |
| DeepSeek V4.1 Flash (1017 build) | 60.5% (DeepSeek scaffold) | Cheaper loops for CI fix bots beside KiCad exports |
Evidence from real KiCad projects (not benchmarks)
Published case studies beat synthetic scores for setting expectations.
JLCPCB: Claude Code + kicad-mcp-pro (XL4015 buck, Oct 2026)
JLCPCB documented a 5 A XL4015 converter: 63 × 33 mm, two layers, 0 ERC / 0 DRC after fixes. The AI hit seven distinct issues (net-label short, FB beside SW, wrong edge clearance, duplicated outline, and more). ERC missed a sense-net short; a pin-by-pin netlist diff caught it. The server had no autorouter; Claude wrote a Python grid router. Full walkthrough: JLCPCB KiCad MCP guide.
JLCPCB: GPT-6 Astra in KiCad (TPA3116 amplifier)
JLCPCB’s Astra case study reports a 44 × 34 mm two-layer board from a four-line brief, native two-sheet schematic, footprint checks against drawings, and 0 ERC / 0 DRC under configured rules (with several rule classes ignored). That is stronger breadth than the buck example on placement plus library hygiene, but the same verification caveats apply. Details: GPT-6 Astra PCB design in KiCad.
OpenAI launch clip (KiCad layout)
OpenAI’s September 2026 launch video shows Astra placing parts and routing from a schematic in under three minutes in a edited clip. Treat it as a first-pass demo, not a fab-ready package without your stackup and test-point rules.
Which model for which PCB task?
| Task | Favor | Why (evidence) |
|---|---|---|
| First KiCad layout from schematic | GPT-6 Astra or Opus 5.5 with Konnect / kicad-mcp-pro | BenchCAD + Astra KiCad case; Opus on long Terminal-Bench |
| Altium live edits and library work | Opus 5.5 or GPT-6.1 Sol + eda-agent | Tool count and script reliability; human approval on each MCP batch |
| High-volume DRC/ERC fix loops in CI | DeepSeek V4.1 Flash (API or self-hosted) | Lower cost per tool call; AutomationBench 60.5% on 1017 build |
| Part selection and architecture only | CELUS, Circuit Mind, or Flux | Purpose-built; stops before fab-critical layout |
| Autonomous placement/routing only | Quilter, DeepPCB, Allegro X AI | Physics engines; engineer pre-routes critical nets |
Verification gates AI still fails without you
- Netlist intent: Export after AI edits; compare every net pin-by-pin. JLCPCB caught ISENSE/ISET with ERC only warning.
- Footprints: Pin 1, courtyard, thermal pads. AI footprint collisions caused 50 association mismatches in JLCPCB’s EasyEDA Astra write-up.
- SI/PI: DRC-clean is not impedance-safe. Route DDR and RF yourself; keep AI on housekeeping nets.
- DFM: Match trace/space and drill rules to your contract manufacturer capability sheet, not only KiCad defaults.
- Supply chain: Query LCSC/JLC parts directly; MCP part search returned bad matches in the buck study.
Security and workflow hygiene
- Run MCP over stdio on a dev machine; review server code for network exfiltration.
- Use a git sandbox per AI project; kicad-mcp-pro’s read-only default until you trust the prompt.
- Enable KiCad API once; keep a single KiCad instance (second window steals the API).
- Log tool calls, not only chat, when models hide reasoning traces (Astra system card note).
How Techwall fits OEM PCB programs
We are a B2B design and manufacturing house. Agents can draft layout faster; we still own EVT boards, firmware, test fixtures, and volume PCBA. Use AI for netlist entry, placement experiments, and DRC cleanup. Send us the git diff, BOM, and constraint document for DFM review before NPI.
If you are comparing Opus 5.5, GPT-6 Astra, and DeepSeek V4.1 Flash for an internal agent pilot, start with read-only MCP on a sandbox board, measure issues per hour, then enable writes. Request a consultation when you need production sign-off, not another demo.
FAQ
Can ChatGPT or Claude design a PCB by themselves?
Only through tools. They need MCP, computer use, or a native AI EDA product. Output always needs engineer verification.
Which KiCad MCP server should we start with?
For safety, kicad-mcp-pro in read-only review mode. For maximum KiCad 10 IPC integration, Konnect. For analysis-only CI, Seeed’s server.
Does Altium have built-in generative layout AI?
Altium documents AI around requirements and cloud assistants, not a single generative autorouter brand. Many teams pair Altium with Quilter, DeepPCB, or MCP agents.
Is GPT-6 Astra better than Opus 5.5 for PCB work?
Astra leads BenchCAD and published KiCad demos. Opus 5.5 leads several long-horizon agent indexes. Run both on the same sandbox netlist and count fixes needed.
Is DeepSeek V4.1 Flash good enough for PCB agents?
Strong for cost-sensitive automation and repo-style fixes per DeepSeek’s AutomationBench rows. It trails Opus on the longest Terminal-Bench horizons.
Is zero DRC enough to order boards?
No. DRC checks your rules file, not design intent, SI, or fab limits. JLCPCB and Altium both state qualified review is still required.
Related on Techwall: GPT-6 Astra charts · DeepSeek V4.1 Flash benchmarks · Robotics components OEM · Contact
Sources: OpenAI GPT-6 Astra launch; DeepSeek V4.1 Flash model card and API changelog; Artificial Analysis; JLCPCB KiCad MCP and Astra KiCad case studies (Oct–Nov 2026); Adafruit Konnect post; GitHub READMEs for Konnect, kicad-mcp-pro, eda-agent; JLCPCB AI tools guide Sep 2026.
