Tim Gerhard's November 14, 2026 write-up Agentic AI Designs an Analog Synthesizer documents Touchsynth β a mixed-signal board where Claude Opus 5.5 inside Claude Code handled circuit architecture, ngspice benches, KiCad layout scripts, firmware, and SSH-controlled lab bring-up. Hacker News surfaced the story as one of the first public, reproducible examples where an agent crossed schematic β simulation β PCB β audio with only floorplan and touch-key fixes left to a human.
If you build with Claude Code commands daily, Touchsynth is the hardware mirror of Cinecode film skills β agent ownership of multi-file artifacts, not a single README summary. Claude Opus Touchsynth PCB 3D render of agent-designed analog synthesizer board
Image: Tim Gerhard
TL;DR #
| Question | Answer |
|---|---|
| MCU | CH32V003 RISC-V |
| Architecture | Juno-style DCO ,Moog ladder VCF ,VCA , touch keys |
| Tooling | Claude Code (Opus 5.5) ,KiCad 10 ,ngspice ,Freerouting , Python glue |
| Calendar time | ~19 h prompt β PCB order (per author log) |
| Human edits | PCB floorplan ,4-layer routing strategy ,touch sensitivity |
| Primary source | Gerhard blog andTouchSynth GitHub repo |
Why this post matters for AI builders #
Most agentic coding benchmarks stop at LeetCode or web apps. Touchsynth forces physical constraints:
- JLCPCB free parts cap op-amp and capacitor choices β the agent could not spec a fantasy OTA.
- ngspice either passes orsmokes the ladder filter β no hiding behind TypeScript mocks.
- DRC and routing fail loudly when analog blocks crowd digital touch sensors.
That is the same maturity jump Edward Witten lattice paper with Opus hints at in math, but here the proof is audio on a scope.
Claude Opus Touchsynth agent flow covers analog PCB design from ngspice to KiCad routing
Gerhard methodology: spec-first, not patch-first #
Gerhard states he follows a V-model discipline:
- Let the agent propose concepts from the JLCPCB parts CSV (1,586 lines pulled via JSON API).
- Lock a written specification (Juno DCO + ladder VCF).
- Generate ngspice test benchesS1βS7 and asimulation report .
- Build a Python behavioral model , calibrate against SPICE, renderaudio demos .
- Only then export KiCad sheets andPython placers +Freerouting .
He argues iterative netlist hacking fails because context windows drown in partial edits β a lesson that transfers to loop engineering for coding agents when you let agents rewrite half a monorepo without tests.
Circuit highlights the agent chose #
Gerhard's review (without changing nets) calls out: DCO integrator β TL072 integrator, 10 nF X7R 50 V cap, LM324 V-I stage, PWM pitch control β chosen for leakage and linearity within free parts.
Transistor ladder filter β 24 dB/oct Moog-style ladder built from matched NPN pairs, tail current exponential converter, instrumentation amp output pickoff β Gerhard links classic Moog ladder explainers for readers who want theory beyond the agent report.
VCA β differential pair with 74HC595 + R-2R trim DAC because PWM outputs were exhausted after filter and resonance control.
Touch keys β CD4051 muxes and 74HC595 chains because CH32V003 pins ran out β classic agent pin-budget story.
Layout: where humans still steer agents #
Gerhard's largest intervention was PCB floorplanning:
- Initial agent placement bunched analog blocks ; routing failed ontwo layers .
- Switching to four layers plusfunctional-block floorplan unlockedFreerouting .
- Gerhard moved blocks to open routing channels betweenanalog sensitive sections anddigital touch strips.
That matches how Claude bughunter skills still need human threat models β agents propose, humans enforce separation of concerns.
Bring-up: agents closing the loop on real hardware #
After JLCPCB assembly (~two weeks), Gerhard:
- Connected a lab PC over SSH to Claude Code.
- Fed audio loopback into the MCUADC .
- Read Rigol MSO5074 traces over Ethernet while the agent stepped calibration scripts.
SPICE benches S1βS4 matched Board1 captures closely β Gerhard notes surprise because second-source transistor models often diverge.
Only fixes: manual MCU solder bridge removal and touch recalibration (keys too hot).
Demo tracks on the project site are sheet-music-driven, multitracked by recording one voice at a time β honest about monophonic hardware limits.
Touchsynth mixed-signal PCB design with Claude Code spans simulation validation and automated oscilloscope bring-up
What developers are saying on Hacker News #
The Hacker News thread was thin at fetch time (single early comment), so there is no substantive What developers are saying section yet. Worth revisiting if hands-on builders post distortion measurements or BOM cost breakdowns.
Limits and safety notes #
- Analog design agents can stillover-componentize when the parts library is narrow β Gerhard notesexcess Moog-era circuitry becauseno OTA/JFET sat in the free list.
- Touch keys on bare PCB arehard to play ; Gerhard drives viaMCU scripts instead of performance MIDI.
- Export and ITAR are not in scope β this ishobby/education scale, not defense electronics.
- Do not point agents atexport-controlled EDA flows without legal review; Touchsynth staysopen BOM + JLC .
Toolchain specifics worth copying #
Gerhard's post lists concrete versions β rare in hype threads:
- KiCad 10 on Windows withWSL for Linux CLI tools.
- ngspice test benches namedS1βS7 with downloadable reports.
- Freerouting attempted ontwo layers , abandoned untilfour-layer stack-up.
- CH32V003 firmware built with theRISC-V toolchain Gerhard already uses on other projects.
The 19-hour log table is a realistic agent calendar cost: most hours were unattended autorouters, not token streaming. Budget similar wall time if you replicate β and compare that to human-only ladder-filter weeks from analog forums.
Audio demos came from the behavioral Python model before boards arrived β exactly the digital twin pattern Sophon realtime voice pursues in software, but here grounding is SPICE, not mel spectrograms.
Safety, export, and lab discipline #
Touchsynth is bench-scale, but teams in defense-adjacent EDA should note:
- Automated SSH scope control is powerful β treat lab PCs likeCI runners withnetwork segmentation .
- Agent-written firmware still needsmanual power-on smoke tests beforeloudspeakers face humans.
- JLCPCB assembly isnot ITAR-aware ; this BOM isconsumer hobby , notmunitions .
For software-only readers, the moral is identical to Claude desktop Linux telemetry fixes: agents touch real side effects once they leave the chat box.
How to try the pattern yourself #
- Mirror Gerhard repo structure β separate
spec/,sim/,pcb/,fw/folders so Claude Code can resume deterministically. - Force simulation gates before layout β refuse KiCad export untilngspice reports attach to the commit.
- Give the agent instrument APIs β SSH to scope and programmer like**connect fitness MCP patterns** but forlab gear .
- Budget human hours for floorplan β agents still missreturn current paths on two-layer budgets.
Compare to NVIDIA SkillEvaluator for agent skills β different domain, same meta-lesson: evaluate end-to-end outcomes, not single-file diffs.
BOM economics and reproducibility #
Gerhard constrained parts to JLCPCB's free list β a contrived but repeatable teaching lab. Readers replicating should expect:
- BOM cost intens of dollars at hobby scale,plus****MCU programmer andscope .
- Four-layer boards beattwo-layer for mixed signal, butraise****Fab minimums β the agent only succeeded after layer bump.
- Sound examples on the project site werepre-silicon frombehavioral models β audition those before ordering boards to hearfilter sweeps .
Compare total calendar time (~19 h agent + two-week fab) to hiring a consulting EE for ladder filter tuning; the post's value is documentation density, not guaranteed cheaper hardware.
Failure modes Gerhard does not glamorize #
The blog is candid about touch keys proving unplayable at human scale and about excessive analog circuitry because JLC lacked OTAs. Those are the wrinkles marketing threads skip:
- Agents over-spec passives whenpin mux runs out β watchBOM creep .
- Autorouters burn hours withouthuman floorplan hints β budgetcalendar , not justtokens .
- SSH scope control candamage gear if scriptsclip supply rails β keepmanual estops .
Treat Touchsynth as R&D lab notes, not a product roadmap for music hardware startups.
If you file Issues against Claude Code hardware skills, cite Gerhard's V-model commit boundaries β future maintainers will thank you when KiCad netlists stop drifting from SPICE goldens. Handoff tip: archive ngspice log hashes beside Gerber ZIPs in release tags so the next agent session cannot claim DRC clean without reproducing S3 filter gain sweeps first. Hardware CI is sparse today; treat these artifacts as your regression suite until open EDA agents mature. Screenshot scope FFTs too β audio regressions hide in harmonics, not Git diff stats.
Related on explainx.ai #
Hardware claims follow Gerhard November 14, 2026 primary post. Audio and BOM live in the linked project repository.