Multisim MCP

Unofficial Windows MCP server for creating NI Multisim schematics, running experiments, exporting data, and generating reproducible reports.

非官方 Multisim 自动化 MCP：从受限 SPICE 网表生成可编辑电路图，调用本机 Multisim 执行实验，并导出 .ms14、原理图、raw、CSV、SVG 和 Markdown 报告。

MCP Core 1.2 release candidate; the current public stable release is 1.1.0. This source/package does not include the React Workbench frontend. It may expose optional loopback bridge APIs for compatible local clients. This project is not affiliated with NI. Multisim must be installed and licensed locally. COM runs in an isolated 32-bit Python worker; the MCP frontend may use 32-bit or 64-bit Python.

Linux and Docker do not run Multisim COM. With a local ngspice installation they can run safe SPICE simulation, complete experiments, verification, and the same optimization/correction services; otherwise they retain MCP initialization, tool discovery, and runtime_status diagnostics. COM-backed operations still fail closed outside Windows. The minimal non-root image contains no EDA runtime, NI software, samples, licenses, or extracted templates.

Current capability

Stable and verified on Multisim 14.3:

• MCP stdio lifecycle and isolated worker/frontend runtime diagnostics.
• Open/save circuits and enumerate components, inputs, and outputs.
• DC operating point, AC sweep, single-frequency AC, and transient analysis.
• Input waveform injection and RLC value read/write.
• SPICE netlist execution with safe op, dc, ac, and tran commands.
• Netlist, BOM, schematic image, raw data, CSV, SVG, Markdown, standalone bilingual HTML/PDF, and reproducibility-manifest export.
• High-level run_circuit_experiment workflow.
• Read-only plan_design_options / select_design_option / prepare_design_specification / prepare_netlist_draft / resolve_component_requirements / approve_component_resolution / compile_executable_netlist / approve_executable_netlist workflow that compares 2--4 technical paths, binds the selected option, collects electrical parameters, then emits an explicitly approved but non-executable logical block network and a separate human-reviewed component approval artifact. The first bounded compiler supports only signal-passive and emits an in-memory pin-level CircuitDesign/SPICE preview; the separate approve_executable_netlist gate binds the preview digest and only opens schematic planning; file writes, stimuli, analysis, and simulation remain later explicit gates.
• Durable submit_circuit_experiment queue with progress, cancellation, total/heartbeat timeouts, and isolated-worker recovery.
• Durable submit_design_optimization with candidate-level checkpoints, evidence revalidation, and non-overwriting interrupted-attempt recovery.
• Deterministic read-only diagnose_design for topology, requirement, convergence, and evidence-backed BJT/op-amp saturation findings.
• Evidence-backed evaluate_design_patch for unchanged-baseline versus one explicit in-memory candidate, with before/after diagnoses and no auto-adoption.
• Mixed topology/value global_optimize_design with exhaustive or deterministic Halton search, hard constraints, epsilon-aware Pareto fronts, and no auto-write.
• Bounded model-planned autonomous_correct_design, where every topology/value proposal must compile and pass real experiment gates before it can advance.
• Durable submit_global_optimization and submit_autonomous_correction jobs with integrity-checked candidate/round recovery and no persisted API keys.
• benchmark-suite offline/real gates for RC, RLC, op-amp, BJT, and regulated power-supply correction; the local 2026-08-25 real gate passed all five cases.
• course-demo builds the bilingual five-waveform course-design contract with 12 explicit verification gates, a 35-row BOM, and a model-evidence readiness gate, then can run the same contract through the selected Multisim/ngspice backend. The bundled fixture is explicitly behavioral; a component-level claim requires HE555/74LS74/LM324/1N4007 provenance plus an integrity-checked real Multisim 12/12 result.
• A real Windows workstation gate has exercised Multisim 14.3 COM, ngspice 47, the DFF behavioral reference, the five-waveform 12/12 contract, and an RC cross-backend comparison. See REAL_RUNTIME_VALIDATION for reproducible commands and evidence boundaries.
• inspect-project builds a bounded, read-only project snapshot from versioned directory manifests for compatible local clients.
• execute-handoff validates a downloaded controlled-execution handoff and, only with explicit --confirm, runs its schematic-first/verified-simulation sequence without bypassing approval or manifest gates.
• Versioned ExperimentSpec verification with persisted PASS/FAIL/unverified evidence and theory-versus-simulation error.
• Parameter, tolerance, temperature, and seeded Monte Carlo sweeps with a 100-run hard limit and durable-worker support.
• MCP 2026-07-28 discovery plus automatic compatibility with legacy clients.
• Seventeen experiment artifact Resources (including the SPICE compatibility audit), two sweep Resources, one job-status Resource, and five bilingual workflow Prompts.
• Validated structured output for the complete experiment workflow.

Completed experiments return an opaque experiment_id and resource URIs such as:

multisim://experiments/{experiment_id}/manifest
multisim://experiments/{experiment_id}/report
multisim://experiments/{experiment_id}/schematic
multisim://experiments/{experiment_id}/data
multisim://experiments/{experiment_id}/plot
multisim://experiments/{experiment_id}/verification
multisim://experiments/{experiment_id}/spice-compatibility
multisim://experiments/{experiment_id}/formal-html-zh
multisim://experiments/{experiment_id}/formal-html-en
multisim://experiments/{experiment_id}/formal-pdf-zh
multisim://experiments/{experiment_id}/formal-pdf-en
multisim://experiments/{experiment_id}/reproducibility-manifest
multisim://sweeps/{sweep_id}/summary
multisim://sweeps/{sweep_id}/data

Completed durable jobs restore their handles after restarting the server. For other historical output directories, call register_experiment_artifacts to restore a process-local handle. Resource reads are limited to the fixed artifact set and default to 16 MiB per file; set MULTISIM_MCP_RESOURCE_MAX_BYTES to a positive integer to change that limit.

Experimental:

• Editable schematic generation supports R/L/C, scalar and waveform voltage/current sources, B/E/F/G/H controlled sources, T/O/U distributed lines, coupled inductors, modeled diodes, NPN/PNP BJT, NMOS/PMOS, JFET/MESFET, voltage switches, five-terminal op-amps, and generic two-to-sixteen-terminal X subcircuits. Extended families currently use verified carrier symbols pending dedicated artwork. Compatible inline vendor .subckt models are recursively expanded into editable primitives with nested dependencies and instance parameters retained. The result reports editable_model_coverage; conditional/proprietary records that cannot be expanded remain explicit carrier-only evidence. Native NOT/AND/OR/NAND/NOR/XOR/XNOR and JK flip-flop symbols are available as a preview; their open/export and authoritative timing-data paths are verified. Native XSC oscilloscope and configurable XFG function-generator state are generated alongside authoritative CSV/SVG experiment data. Multisim's exported native netlist is checked after opening so silently omitted parts fail the run. The user-local component-pack workflow also supports a verified TIMER8 carrier for LM555CN and a DFF8 A-section carrier for 7474N; portable XU1 instances normalize to native U1. Vendor timer/digital bodies may be omitted by ReportNetlist, so native component enumeration is recorded as the authoritative presence evidence for these carriers. DFF8 is a functional 74LS74 substitute, not an exact 74LS74N/74LS74D model claim. Structured correction/optimization rebuilds keep both carriers as strict eight-terminal X... records. Use the isolated tools/probe_native_replacement.py helper for crash-safe replacement tests.
• Portable @KIND adapters synthesize transformer, potentiometer, relay, crystal, power semiconductor, D/T flip-flop, four-bit counter/register, and one-bit ADC/DAC models from ordinary primitives. See docs/COMPONENT_ADAPTERS.md.
• Generated schematic probes are not enabled by default. Experiment data is obtained authoritatively from the same netlist through Multisim's engine.

Platformization work now includes strict versioned CircuitDesign, DesignPatch, and ArtifactSet objects, the EdaBackend protocol, a transport-neutral application service, and an injectable Multisim adapter. See docs/EDA_CORE.md. The current MCP tools retain their 1.0 signatures while execution is migrated behind this boundary.

Install

Requirements:

• Windows and a licensed Multisim 14+ installation.
• Python 3.10+ for the MCP frontend (32-bit or 64-bit).
• A separate 32-bit Python 3.10+ environment containing this package and pywin32 for the Multisim worker.
• Node.js 18+ only for .ms14 XML conversion.

The Windows dependency set deliberately uses cryptography>=48.0.1,<49. 48.0.1 is the newest release line currently providing an official win32 wheel; newer releases would otherwise make 32-bit installation attempt an unsupported local Rust build. This boundary should be reviewed whenever a new win32 wheel becomes available.

Install the Python package once; the server launcher never installs packages or writes setup logs to MCP stdout:

cd mcp_server
.\setup.ps1 -Python C:\path\to\python32\python.exe
npm install --global electronics-workbench-decoder@0.2.0

The public wheel is intentionally code-only: it contains the provenance manifest but no XML extracted from NI samples. Before generating schematics, build a local component pack from your own licensed installation as described below and set MULTISIM_MCP_TEMPLATE_DIR. Other Automation API tools can still be installed without that pack.

Diagnose the installation before starting the server:

# Human-readable output (Chinese or English)
C:\path\to\python32\Scripts\multisim-mcp.exe doctor --lang zh
C:\path\to\python32\Scripts\multisim-mcp.exe doctor --lang en

# Explicitly start/connect to Multisim and verify licensing plus COM activation.
C:\path\to\python32\Scripts\multisim-mcp.exe doctor --connect

# Stable JSON for agents and CI. Add --strict to require the full workflow.
C:\path\to\python32\Scripts\multisim-mcp.exe --json doctor
C:\path\to\python32\Scripts\multisim-mcp.exe doctor --json --strict

doctor is side-effect free by default: it does not activate Multisim. It checks the Python version and architecture, pywin32, the 32-bit COM registration, the local template pack, and the pinned .ms14 codecs. A normal diagnostic run returns exit code zero even when setup is incomplete so an agent can parse all checks. --strict returns non-zero unless the complete workflow is ready. --connect is the explicit opt-in that may start Multisim; it restores a previously disconnected COM state after the probe and does not disturb an already connected instance.

Build a bounded, read-only project snapshot for a compatible local client:

C:\path\to\python32\Scripts\multisim-mcp.exe inspect-project `
  --root C:\msre_exp --json

The command verifies manifest-referenced artifacts by default, reports corrupt child directories without modifying them, and never treats an unmanifested directory as an experiment. See docs/PROJECT_INSPECTION.md.

For a compatible local control surface, start the optional read-only bridge in a second terminal. It binds to loopback, fixes the project root at startup, and exposes the bounded project snapshot plus opaque-handle experiment detail/media and optimization evidence views:

$env:PYTHONPATH = (Resolve-Path .).Path
python -m multisim_mcp.cli workbench-api --root C:\msre_exp --port 8787

The client can read that bounded snapshot. Experiment details reuse the existing Resource summary and media must be referenced by a verified directory manifest. The Models / API page also reads the secret-free provider metadata from /api/provider-config and can explicitly probe one OpenAI-compatible /models endpoint through /api/provider-probe; it never accepts plaintext credentials and does not write provider files. Use the copied multisim-mcp configure ... --apply command for persistence. Do not bind this bridge to a public interface; it is not a remote or multi-tenant API.

The planning page uses the same loopback bridge for the read-only design flow: /api/design-plan, /api/design-plan/select, /api/design-specification, /api/netlist-draft, /api/component-resolution, /api/component-resolution/approve, /api/executable-netlist/compile, and /api/executable-netlist/approve. After the netlist review, /api/executable-netlist/simulation-approve binds a safe ExperimentSpec to the approved preview. These routes only return bounded previews or approval artifacts; schematic generation and simulation remain explicit MCP/CLI operations.

To execute the two-step handoff without manually copying each MCP call, first run the validation-only command from a trusted terminal:

python -m multisim_mcp.cli execute-handoff `
  --handoff .\multisim-approved-experiment-handoff.json `
  --root C:\msre_exp --json

After reviewing the resolved output directory and approval identities, add --confirm to execute schematic generation and then the verified experiment. The CLI rejects absolute/traversal paths, mismatched netlists or approval payloads, and existing artifacts are not overwritten unless the handoff requests it and the operator also supplies --allow-overwrite. See CONTROLLED_HANDOFF. For long experiments, add --submit --confirm instead; the schematic is generated first and the verified experiment is then queued for the durable worker. Its job handle is available through the MCP/CLI job-status surfaces.

The Jobs / 队列 page polls /api/jobs every four seconds and shows only sanitized durable-job state (queued, running, succeeded, or failure class). /api/jobs/{job_id} exposes the same bounded status for one opaque handle; specs, result payloads, output paths, logs, and error text stay behind MCP/CLI surfaces. The page has no submit, cancel, retry, or other mutation controls. For a succeeded job, the single-job endpoint resolves an opaque result_entry only when the output directory is an exact, non-symlink child of the fixed project root and its experiment/optimization manifest passes the configured integrity check. The browser then refreshes the snapshot before opening it.

Optimization details are also manifest-backed and bounded. They expose run state, budget, ranked candidates, objective convergence, Pareto layers, and the recommended solution, plus a descriptive observed-candidate-range sensitivity summary and a bounded, read-only next-search proposal. Numeric proposals use a small E24 neighborhood; categorical proposals repeat observed values. The proposal never starts experiments or mutates the optimization specification. The spec_draft payload is explicitly non-executable and can be copied/downloaded for manual review. Its preflight only checks bounded budget/value limits and the manual-approval requirement; it never issues an approval token or enables execution. The separate local search-plan-approve CLI can issue a short-lived, one-time token after review. The token is bound to the opaque entry handle, optimization ID/kind, normalized source-design SHA-256, source-spec SHA-256, complete canonical draft SHA-256, and budget summary; search-plan-verify only validates and does not consume or execute anything. After review, search-plan-submit revalidates those exact inputs, consumes the token, and queues one derived bounded optimization or global_optimization job. It is a durable queue hand-off (execution_started=false), so a long-lived MCP worker using the same MULTISIM_MCP_JOB_DIR performs the actual experiments. The queue record retains approval/binding digests but never the bearer token; approval-bound replay is idempotent after a crash. Topology-choice drafts must use explicit topology operations rather than scalar value substitution. They omit patch paths, experiment directories, and raw error paths; sensitivity is not a causal derivative or a global-optimality proof.

Start the server. Calling multisim-mcp without a subcommand remains backward compatible; serve is the explicit equivalent:

.\run_server.ps1
C:\path\to\python32\Scripts\multisim-mcp.exe serve

For a 64-bit frontend, install the package in both environments and either let the Windows py launcher discover the worker or select it explicitly:

$env:MULTISIM_MCP_WORKER_PYTHON = 'C:\path\to\python32\python.exe'
C:\path\to\python64\python.exe -m multisim_mcp.server

Generate a client configuration fragment:

# Claude Desktop JSON
C:\path\to\python32\Scripts\multisim-mcp.exe config `
  --client claude-desktop `
  --python C:\path\to\python32\python.exe `
  --template-dir C:\MultisimMcp\component-pack `
  --work-dir C:\msre_exp

# Codex config.toml
C:\path\to\python32\Scripts\multisim-mcp.exe config `
  --client codex `
  --python C:\path\to\python64\python.exe `
  --worker-python C:\path\to\python32\python.exe `
  --template-dir C:\MultisimMcp\component-pack

# Official DeepSeek Harness Cordis plugin row
C:\path\to\python32\Scripts\multisim-mcp.exe config `
  --client deepseek-harness `
  --python C:\path\to\python32\python.exe `
  --template-dir C:\MultisimMcp\component-pack `
  --work-dir C:\msre_exp `
  --artifact-export-dir C:\MultisimMcp\exports `
  --tool-profile experiment

# Install the five bilingual workflow skills in a Harness project.
C:\path\to\python32\Scripts\multisim-mcp.exe harness-skills --output .dsh/skills

# Safely discover model-provider environment settings (preview only).
C:\path\to\python32\Scripts\multisim-mcp.exe configure --auto --json

# Atomically store references after reviewing them.
C:\path\to\python32\Scripts\multisim-mcp.exe configure --auto --apply

# One explicit tool-free model request from a UTF-8 file.
C:\path\to\python32\Scripts\multisim-mcp.exe model --input .\prompt.txt --json

# Four design tools plus four optional completed-experiment evidence tools.
C:\path\to\python32\Scripts\multisim-mcp.exe model-diagnose `
  --input .\diagnosis-prompt.txt --netlist .\circuit.cir `
  --experiment-dir .\completed-experiment `
  --enable-patch-preview `
  --audit-output .\agent-audit.json --json

# Unwrapped command/args/env JSON for another stdio client
C:\path\to\python32\Scripts\multisim-mcp.exe config --client generic

The generator previews content on stdout. --output <new-file> writes a fragment, refuses to overwrite by default, and accepts --force only when the caller explicitly wants replacement. It does not merge into a live client configuration.

The Harness fragment uses @deepseek-ai/dsh-mcp-client, enforces the upstream 1-32 character serverName rule, and never copies a DeepSeek API key into the MCP child process. See docs/DEEPSEEK_HARNESS.md. The same --tool-profile core|experiment|optimization|full option works for every generated client config. The default remains full for compatibility. --artifact-export-dir sets the only root beneath which the artifact export tool may write; without it, artifact export fails closed. harness-skills writes the packaged bundle to .dsh/skills, refuses existing files by default, and only replaces them when --force is explicit. Source-tree maintainers can run python tools/check_deepseek_harness_compat.py --json from the repository root to validate the pinned Harness contract. The independently installable Harness bundle lives in integrations/deepseek-harness and is published as multisim-mcp-dsh-plugin@1.1.0. Install it with dsh plugin --profile web add multisim-mcp-dsh-plugin@1.1.0. The separate configure command prepares model-provider settings for compatible clients. It never copies credential values into its versioned JSON file or the MCP child process, and it performs network I/O only with explicit --probe. See docs/MODEL_PROVIDER_CONFIGURATION.md. The transport-neutral runtime adds bounded non-streaming Chat Completions, normalized usage, cooperative cancellation, double-opt-in failover, and an allowlisted library-level tool loop. The ordinary model command deliberately exposes no tools and accepts no inline prompt arguments. The separate model-diagnose command explicitly enables four read-only tools over strict CircuitDesign JSON or safely parsed SPICE, without starting a backend. Explicit --experiment-dir adds four more read-only tools over a sanitized completed-experiment snapshot: waveform-column statistics, requirement verdicts, and artifact hashes. It excludes report text, raw samples, artifact content, and paths, and marks the design/experiment association as unverified. Explicit --enable-patch-preview adds one non-persistent tool that validates a bounded DesignPatch, derives its inverse, and returns structural deltas. It does not mutate the design, write files, call a backend, simulate, or approve. Opt-in --audit-output writes model rounds, validated tool calls, hashes and usage to versioned JSON while excluding prompts, answers, reasoning, credentials, and full tool results; replacement requires --audit-overwrite. The separate local patch-approve, patch-apply, patch-revert, and patch-recover commands provide short-lived one-time approval, durable crash journaling, and verified commit/rollback for CircuitDesign JSON. The additional patch-verify-approve, patch-verify-apply, and patch-verify-recover commands bind an explicit verification plan and complete experiment evidence, simulate the in-memory candidate through Multisim, and persist only an all-pass verdict. These mutation commands are not exposed to the model. The diagnose-design CLI and diagnose_design MCP tool run without a model, COM activation, simulation, or writes. A completed experiment may be attached only when its recursive manifest verifies and its canonical netlist matches the input design. See docs/DESIGN_DIAGNOSIS.md. The evaluate-design-patch CLI and evaluate_design_patch MCP tool run the unchanged baseline and one explicit reversible candidate under the same hard requirements. They retain both experiments, the inverse patch, before/after diagnoses, and a recursive manifest; even an all-pass candidate requires a separate approval before persistence. See docs/DESIGN_PATCH_EVALUATION.md. The separate optimize-design CLI and optimize_design MCP tool evaluate explicit or E12/E24/E48/E96 scalar component values under a hard budget (32 experiments including the baseline). Electrical requirements, optional in-stock rules, and variable-cost ceilings are hard constraints; inventory binds values to part/supplier/cost evidence, and cost only breaks equal objective ties. The source design is never modified, and the selected patch still requires the verified approval workflow. Each patch, experiment, objective, procurement verdict, failure, stopping reason, CSV comparison, and recursive manifest is retained. See docs/DESIGN_OPTIMIZATION.md. Mixed finite-domain topology/value search is documented in docs/GLOBAL_OPTIMIZATION.md. The bounded diagnose-propose-simulate-select loop is documented in docs/AUTONOMOUS_CORRECTION.md. Neither workflow persists a selected patch automatically. Long searches can use submit_design_optimization, then the existing get_experiment_job, cancel_experiment_job, and retry_experiment_job tools. Completed candidate evidence is verified before reuse; an uncommitted candidate is rerun in a new attempt directory. The local equivalent is optimize-design --resume against a matching interrupted directory. The compare-designs CLI and compare_design_variants MCP tool run one verification plan across 2–16 complete designs, rank only finite measured all-pass variants, and retain every design, experiment, error, CSV rank, and recursive manifest without changing or adopting an input. See docs/DESIGN_COMPARISON.md. See docs/MODEL_PROVIDER_RUNTIME.md and docs/READ_ONLY_EDA_DIAGNOSIS.md, plus docs/DESIGN_PATCH_TRANSACTIONS.md.

Manual MCP client configuration:

{
  "mcpServers": {
    "multisim": {
      "command": "C:\\path\\to\\python32\\python.exe",
      "args": ["-m", "multisim_mcp.server"]
    }
  }
}

Call runtime_status first when diagnosing installation problems.

CLI JSON contract

multisim-mcp --json doctor emits one JSON object with schema_version, success, readiness booleans (including activation state), runtime facts, and a stable checks[] array. Each check has a stable id, status, and message; failed checks may include repair. Missing setup is reported as data rather than a JSON error.

multisim-mcp config ... --json emits a result envelope containing the client, server name, optional output path, and generated content. Invalid input or an overwrite refusal uses this error shape and exit code 2:

{
  "schema_version": 1,
  "command": "config",
  "success": false,
  "error": {"type": "ValueError", "message": "..."}
}

JSON stdout never contains progress text or credentials. Import-time COM cache diagnostics are redirected to stderr.

User-local component packs

To keep licensed/reverse-engineered component assets separate from the open engine, a contributor can derive a local pack from the NI samples installed on their own machine:

$env:PYTHONPATH = (Resolve-Path .\mcp_server).Path
C:\path\to\python32\python.exe .\tools\bootstrap_local_component_pack.py `
  --output C:\MultisimMcp\component-pack
$env:MULTISIM_MCP_TEMPLATE_DIR = 'C:\MultisimMcp\component-pack'
$env:MULTISIM_MCP_TEMPLATE_ONLY = 'true'

For a locally verified native model, overlay the component into that user-local pack instead of copying NI assets into the repository:

python .\tools\overlay_local_component_pack.py `
  --pack C:\MultisimMcp\component-pack `
  --source C:\path\to\verified-lm324.ms14 `
  --refdes U5 --kind OPAMP5 --identity-token LM324M --force

The tool records source and installed-template hashes, backs up replaced files, and marks the manifest as local-only. It never publishes the .ms14, decoded XML, or vendor model body. Use the actual saved reference designator from the probe project; it may differ from the design's intended U1/D1 labels.

The generator connects to licensed Multisim and creates a temporary blank circuit so the project shell matches the installed file-format version. Save open work first. Version 1.0 writes a schema-2 local manifest; doctor rejects schema-1 packs generated by alpha releases and asks the user to rebuild them.

The configured pack is the public release's schematic-template source. A local development checkout may contain ignored fallback templates, but public wheels do not. The generated manifest records relative sample provenance. Local reverse-engineering authorization does not itself grant permission to publish the resulting XML files.

Recommended agent workflow

Before generating a netlist, call plan_design_options with the functional requirements, hard constraints, context, and objective weights. Present the bounded options and the default recommendation to the user, wait for an explicit plan_id + option_id selection, then call select_design_option with the original envelope, then call prepare_design_specification until all required electrical parameters are present. After the user reviews that exact specification digest, call prepare_netlist_draft with explicit approval to inspect the logical module/net structure and unresolved component requirements. Then call resolve_component_requirements; after supplying explicit candidates, ratings, and model provenance, call approve_component_resolution to bind the human review gate. The approval artifact authorizes only a later compiler and does not itself generate SPICE. For an option in the explicit support matrix, call compile_executable_netlist; it revalidates the artifact and local model hashes, and currently supports only signal-passive. Then call approve_executable_netlist with explicit review confirmations; its approval only opens schematic planning and remains unapproved for file writes or simulation. To consume that approval, pass the complete preview as executable_netlist, the approval artifact as netlist_approval, and the preview's exact spice_netlist to create_schematic_from_netlist. The existing schematic tool revalidates both artifacts before writing .ms14, rejects a missing or mismatched handoff, records the approval identifiers in its result, and still never starts simulation. Next call approve_simulation_plan to bind the same preview and netlist approval to a validated ExperimentSpec; pass those three artifacts to run_verified_circuit_experiment so it rechecks the commands, measurements, and limits before creating the schematic or starting Multisim. For a durable long-running job, pass the same three artifacts and reviewed requirements to submit_circuit_experiment; they are persisted with the job and revalidated inside its isolated worker before execution. The planner is deterministic and planning-only: declared constraints are retained for the next validation step but are not electrically enforced yet; its scores are heuristic, every option is marked planning-only, and its execution boundary keeps schematic generation, simulation, and file writes false. The logical draft also keeps CircuitDesign, SPICE generation, and schematic readiness false. See DESIGN_PLANNING.md.

Choose the open complete-experiment backend when ngspice is installed:

export MULTISIM_MCP_EXPERIMENT_BACKEND=ngspice
# Optional when ngspice is not on PATH:
export MULTISIM_MCP_NGSPICE=/absolute/path/to/ngspice

Multisim remains the default. run_spice_netlist also accepts an explicit backend argument. Complete ngspice runs publish an honest non-editable schematic.svg connectivity diagram plus PNG preview, not a fake .ms14. See docs/OPEN_EDA_BACKENDS.md. Every new experiment also writes spice-compatibility.json, which records source and executed-netlist hashes, model provenance, dialect features, backend risk, and solver-version evidence. Use audit_spice_compatibility before a run and read the artifact through multisim://experiments/{id}/spice-compatibility. Digital-device runs additionally publish digital-observation.json and expose it through the digital-observation experiment resource. It records whether each digital output was observed, came from a behavioral reference, or remained unobserved. When Multisim omits a digital output, the artifact recommends an explicit ngspice rerun; backend switching is never automatic.

For the supported native D flip-flop carriers, call build_behavioral_reference to produce an explicit, machine-readable reference netlist:

XU1 d pr clr clk q nq 0 vcc 7474N

becomes an @DFF adapter with the documented mapping D, ~PR, ~CLR, CLK, Q, ~Q, GND, VCC → D, CLK, SET, RESET, Q, QBAR, HIGH, LOW; two explicit NOT devices convert the active-low preset/clear inputs to XSPICE's asserted-high set/reset inputs. The returned netlist must be passed explicitly to run_spice_netlist(backend="ngspice"); it remains behavioral evidence and does not establish native 74LS74 timing or electrical equivalence. The ngspice backend compiles this adapter through the same bounded simulation translator used by the Multisim command path, including the required XSPICE bridge and d_jkff model definitions.

For a one-call reference experiment, use run_behavioral_reference with the native netlist and a safe analysis command. It performs the same conversion, forces the explicit ngspice backend, and returns the measured result together with reference_netlist and behavioral_reference metadata. It rejects a netlist with no supported native DFF carrier so that a no-op conversion cannot be mistaken for a reference run. The source netlist is never modified. A local ngspice executable is still required; if it is not installed, the tool reports the backend's clear runtime failure rather than silently switching to Multisim.

The same workflow is available from the local CLI:

multisim-mcp behavioral-reference `
  --netlist C:\experiments\native-dff.cir `
  --commands C:\experiments\reference-tran.txt `
  --output C:\experiments\dff-reference `
  --json

The CLI reads both files as bounded UTF-8 input and exposes the same explicit ngspice-only semantics as the MCP tool.

The high-level tool accepts a SPICE netlist and a safe experiment command:

{
  "netlist": "VIN vin 0 DC 10\nR1 vin vout 1k\nR2 vout 0 1k\n.end\n",
  "commands": "dc VIN 0 10 0.1",
  "output_dir": "C:\\experiments\\divider",
  "title": "Resistor divider",
  "overwrite": false
}

For normal agent use, submit the same arguments with submit_circuit_experiment. It returns immediately:

{
  "success": true,
  "job_id": "job-...",
  "state": "queued",
  "status_uri": "multisim://jobs/job-...",
  "output_dir": "C:\\experiments\\divider"
}

Poll get_experiment_job or the status Resource, and call cancel_experiment_job when needed. A failed, cancelled, or timed-out record can be resubmitted without copying its persisted source through retry_experiment_job. States are queued, running, cancelling, succeeded, failed, cancelled, or timed_out. A successful record contains the same structured result as run_circuit_experiment. list_experiment_jobs omits large results and supports state filtering.

With the default Multisim backend, the synchronous run_circuit_experiment tool will:

1. Validate the supported netlist and analysis command.
2. Generate and encode an editable circuit.ms14.
3. Open the design in Multisim and export schematic.png.
4. Run the requested analysis through Multisim's engine.
5. Export result.raw, data.csv, plot.svg, logs, and report.md.
6. Export Chinese/English HTML/PDF reports, the SPICE compatibility audit, and manifest.json atomically.

With ngspice selected, the same service keeps steps 1 and 4–6, replaces the editable .ms14 step with a labeled SVG/PNG connectivity graph, and records the non-editable backend profile in backend.json. Existing verification, optimization, global optimization, and autonomous correction callers do not change. compare_experiment_backends aligns common signals from two registered runs and reports tolerance-based MAE/RMSE/max-error verdicts; its evidence block also checks that source and executed netlists and model fingerprints match.

Job records are stored as atomic JSON under %LOCALAPPDATA%\multisim-mcp\jobs by default. Set MULTISIM_MCP_JOB_DIR to select another private local state directory. Records contain the source netlist and settings required for restart recovery, so protect and back up that directory according to the sensitivity of your circuit. job_timeout limits the whole workflow; heartbeat_timeout detects a hung Multisim/codec worker. Output publication is guarded by a cross-process sibling lock and remains transactional. A queue-wide lease keeps execution serialized even if multiple MCP frontend processes use the same job state directory.

Design verification and sweeps

Use run_verified_circuit_experiment with an ExperimentSpec containing explicit measurement signals, criteria, tolerances, and optional theoretical values. Supported metrics include scalar statistics, gain, cutoff frequency, bandwidth, rise time, overshoot, ripple, and power. Each requirement is reported as pass, fail, or unverified; unavailable evidence is never guessed.

Use plan_experiment_sweep before execution to preview every rendered run. run_experiment_sweep executes synchronously, while submit_experiment_sweep uses the durable job worker. Modes are parameter, tolerance, temperature, and monte_carlo; Monte Carlo runs use an explicit integer seed. Sweep values are finite numbers substituted into declared {{NAME}} placeholders, and every rendered netlist passes the same safety validator as a normal experiment.

Use create_schematic_from_netlist when only an editable schematic is needed, or run_spice_netlist for netlist-only simulation. For a compiler-approved design, also provide the matching executable_netlist and netlist_approval objects as described above; the schematic tool will use only the preview's bound SPICE text.

Virtual instruments use explicit pseudo-device records in the same netlist:

XFG1 out 0 inv FGEN WAVE=SINE FREQ=1k AMPLITUDE=2 OFFSET=0.5
XSC1 out inv 0 0 out 0 OSCILLOSCOPE

The XSC terminal order is A, B, C, D, EXT+, EXT-. XFG supports WAVE (SINE, SQUARE, or TRIANGLE), FREQ, AMPLITUDE, OFFSET, DUTY, and RISE.

Completed experiment data can also be read through read_virtual_multimeter, analyze_bode_response, and analyze_logic_signals. These tools return structured measurements and edge events; the Bode adapter explicitly leaves phase unavailable when the raw file contains no phase column.

Safety model

• Safe analysis commands are allowlisted: op, dc, ac, and tran.
• do_command_line is disabled by default. It requires the server-side MULTISIM_MCP_ENABLE_UNSAFE_COMMANDS=1 opt-in.
• Runtime npm downloads are disabled. On Windows the npx fallback remains disabled even when opted in because .cmd shims are not safe for caller-controlled paths. Install the pinned codec globally, or set MULTISIM_MCP_EWD and MULTISIM_MCP_EWE to its dist/ewd.js and dist/ewe.js entry points; the server invokes them through node.exe.
• Existing experiment artifacts are not overwritten unless overwrite=true.
• Sweep expansion is capped at 100 runs and only finite numeric substitutions are accepted.
• Asynchronous job specifications are persisted locally for recovery; their state directory must not be shared with untrusted users.
• Optimization resume fails closed if the design, normalized spec, runtime limits, child manifest, verification evidence, objective, or procurement record differs from the persisted checkpoint.
• The server is intended for trusted local stdio clients, not public network exposure. See SECURITY.md in the repository root.

Test

COM-free tests:

$env:PYTHONPATH = (Resolve-Path .\mcp_server).Path
C:\path\to\python32\python.exe -m unittest discover -s mcp_server\tests -p 'test_*.py' -v

Real Multisim E2E:

$env:MULTISIM_MCP_E2E_SAMPLE='C:\path\to\fixture.ms14'
tools\python32\python.exe mcp_server\tests\e2e_mcp_test.py

The E2E test requires a local fixture with RLC components and at least one simulation output; proprietary NI sample designs are not distributed.

License

Project code is MIT licensed. NI Multisim, its formats, trademarks, and locally installed samples remain subject to their respective owners' terms.