ANSYS Mechanical & ACP MCP Servers

A Model Context Protocol (MCP) server that connects any MCP-compatible AI client to ANSYS Mechanical via PyMechanical gRPC. Talk to your simulation in plain English.

The repo also ships a second server, ansys-acp-mcp, for building composite layups in ANSYS Composite PrepPost (ACP) through PyACP. See ACP server.

https://github.com/user-attachments/assets/309952dc-dce0-4aa6-9b06-e8047f32c2de

Compatible clients

Works with any MCP-compatible client, including:

• Claude Desktop

• Cursor

• Windsurf

• VS Code (with Copilot MCP support)

• Continue

Related MCP server: Ansys MCP Server

What it does

Every tool that changes the model names the analysis it used (analysis: index, name, type). With several analyses in the model, analysis_index or analysis_name must be given (index 0 is often not the one meant); expect_analysis_type refuses a mismatch. Set MECHANICAL_MCP_REQUIRE_EXPLICIT_ANALYSIS=0 to restore the old default of index 0.

What ok: true means

Every tool answers with a JSON object that has an ok key. For tools that create something in the model, ok: true means the object was created, scoped and read back - it does not mean the result is valid. A result added before the analysis is solved comes back with max: null, min: null, solved: false and a note (Mechanical itself would report 0). If anything fails after an object was created, the object is deleted again and the response says rolled_back: true.

Requirements

• ANSYS Mechanical 2026 R1 on Windows. Only 2026 R1 (v261) has been exercised; older releases may lack API members used here (the gravity and reaction-force calls differ between releases) and are untested.

• Python 3.10+, ansys-mechanical-core, mcp[cli]>=1,<2

Installation

git clone https://github.com/codersag/mechanical-mcp.git
cd mechanical-mcp
pip install -e .

Optional extras:

pip install -e ".[report]"    # python-docx: generate_report(fmt="docx")
pip install -e ".[postproc]"  # numpy + scipy: linearize_path

MCP Client Configuration

Add the following to your MCP client's config file. The config file location varies by client:

{
  "mcpServers": {
    "mechanical-mcp": {
      "command": "ansys-mechanical-mcp"
    }
  }
}

Or if running from source:

{
  "mcpServers": {
    "mechanical-mcp": {
      "command": "python",
      "args": ["C:/path/to/mechanical-mcp/mechanical_mcp_server.py"]
    }
  }
}

Restart your client after editing the config.

Finding the Mechanical gRPC port

Mechanical exposes a gRPC server whose port you pass to connect_to_mechanical.

In Workbench (Tools → Scripting → Run Script):

port = GetSystem("SYS").GetContainer(ComponentName="Model").StartGrpcServer()
print(port)

If you have multiple Mechanical systems open (e.g. Modal + Static Structural), run the script for each system by replacing "SYS" with the system name (e.g. "SYS 1", "SYS 2").

Usage

1. Open ANSYS Mechanical 2026 R1 and start its gRPC server (see above)

2. Find the port using one of the methods above

3. Connect to your Mechanical instance:

"Connect to Mechanical on port 10000"

Then interact naturally:

"Assign Structural Steel to the body called Plate"

"Generate a mesh with 50mm element size"

"Add a fixed support to the named selection called Base"

"Solve the analysis and get the maximum von Mises stress"

"Convert part Equipment1 to a point mass"

One session at a time

The server holds one connection. Calling connect_to_mechanical with another port replaces it, but only once the new session answers (a dead or wrong port leaves the working connection alone). Only sessions on this machine are supported: the port is probed on 127.0.0.1 first so a dead port fails fast, and a port that accepts connections but never speaks gRPC is detected by a short protocol probe instead of hanging the handshake. The last port is remembered in a per-user folder (%LOCALAPPDATA%\mechanical-mcp, override with MECHANICAL_MCP_STATE_DIR) so a restarted server can reconnect; only the port, a hash of the project folder and the folder's NAME are stored, never a path. A call that fails in transit (timeout, dropped channel) is never re-sent: Mechanical may or may not have run it, so the response says so and you should look at the model before retrying.

Security and configuration

Tool arguments come from an AI agent that may have read untrusted text, so the server is restrictive by default.

Files: network/UNC paths and : (alternate data streams) are refused from the text, only .json (export, linearize) or .docx/.txt (report) are accepted, and an existing file is never replaced unless you pass overwrite=true. export_nodal_results enforces this inside the Mechanical script as well: it writes a uniquely named exclusive temporary file, checks again right before the final rename and refuses (leaving the file) if one appeared meanwhile, and removes only the temporary file it created. Reserved Windows device names (aux.json, COM1.txt, nul, ...) are refused. disconnect_from_mechanical(shutdown=True) force-exits the Mechanical session whoever started it (unsaved work is lost), so it also needs confirm="SHUT DOWN MECHANICAL"; the normal disconnect_from_mechanical() only drops the connection.

Example with an allow-list and scripting enabled:

{
  "mcpServers": {
    "mechanical-mcp": {
      "command": "ansys-mechanical-mcp",
      "env": { "MECHANICAL_MCP_ROOTS": "C:\\Work\\fea", "MECHANICAL_MCP_ENABLE_SCRIPT": "0" }
    }
  }
}

Privacy

The server is meant to be usable on a work machine without telling the AI model, its logs or this repository who you are.

• Replies are redacted (on by default, MECHANICAL_MCP_REDACT=0 switches it off): the user-profile prefix of a path becomes %USERPROFILE%, and your Windows user name, computer name, licence server (port@host), customer / checkout / session / host ids and e-mail addresses are replaced in every tool reply and error. This is pattern-based: a project that has your name in its own file or folder names, or text Mechanical returns in a form the patterns do not know, can still carry it. Use a neutral project folder name if that matters.

• Stored state is minimal. The remembered-session file holds the port, a 16-character hash of the project folder and the folder's name; no path, no user name.

• File access is confined to the folders in MECHANICAL_MCP_ROOTS / ACP_MCP_ROOTS; error messages report how many folders are allowed, never which.

• Nothing leaves the machine. The server talks to Mechanical on 127.0.0.1 and makes no network requests; it does not log tool arguments to a file. Temporary script and output files are removed after each call and stale ones (older than a day) at start-up.

• The repository's .gitignore excludes project files (*.mechdb, *.wbpj), caches, logs and local configuration; tests/test_privacy.py checks the published files for profile paths, e-mail addresses and licence ids.

Geometry selection, nonlinear setup and post-processing

Load and support tools (add_fixed_support, add_frictionless_support, add_displacement, add_force, add_pressure) take a named selection, or "" plus either:

• entity_ids: comma-separated geometry ids, or

• box_mm: "xmin,xmax,ymin,ymax,zmin,zmax" in mm (blank = open side), with entity_type (face, edge, vertex) and optional body_name.

Box coordinates are compared in mm whatever unit the geometry uses (read from GeoData.Unit). A missing or empty named selection, ids that do not exist, or a box that matches nothing is an error and nothing is created. entity_ids may name faces, edges or vertices.

Inspecting the model: list_boundary_conditions (loads and supports with scope and key values), list_contact_regions, list_results and get_result_values (Max/Min with unit once solved), list_named_selections, create_named_selection (from body names or prefix, geometry ids or a box; refuses empty results and duplicate names unless replace=true).

Other model and result tools: set_2d_behavior, set_body_mesh_size, set_nonlinear_settings (changes only the settings you pass), set_contact_type (overwrites every contact region unless limited with region_names / only_if_currently; the response lists each region's previous type), add_contact (two or more bodies), add_normal_stress, add_stress_intensity, export_nodal_results and linearize_path.

Stress linearization flow

1. add_normal_stress (a signed component; linearizing von Mises or stress intensity is refused).

2. solve_analysis (use wait=false + get_solve_status for long solves).

3. export_nodal_results("Normal Stress", "C:/Work/fea/n.json") - writes node ids, xyz_mm (millimetres), values, value_unit, and whether the quantity is signed. Refuses an unsolved model.

4. linearize_path("C:/Work/fea/n.json", "0,0,0", "0,0,12") - a 3D section line, or two numbers in a plane (then the data must lie in one plane, or give slice_mm). Returns membrane, bending (at the start point; the end point has the opposite sign) and peak = total - (membrane + bending), with section_length_mm.

Not verified against a real solved model: steps 3-4 depend on Result.PlotData, which only exists after a solve, and the server's tests never solve. The export script, units and file format are covered by unit tests with a fake object model; the numerical linearization is tested against analytic profiles (pure membrane, pure bending, linear + peak). Treat the first real run with care.

Known limits

• Sign convention of add_pressure is not verified here. The magnitude is passed to Mechanical as given; which way a positive value acts was not checked (it needs a solve and a look at the result), so confirm it in Mechanical before relying on the direction.

• set_contact_type snapshots the type AND the friction coefficient of every region it changes and restores both if any region fails; the response lists the real before/after of both.

• A client-side timeout on solve_analysis marks the solve as possibly still running: the next solve request is refused until you check and pass confirm_not_running=true.

• A finished background solve's result is kept until read (up to five, one hour) and attached to the poll for the same analysis.

• Verified live (Mechanical 2026 R1, an unsolved copy of a project, nothing solved or saved): gravity, loads and supports (named selection / ids / box scopes), remote displacement, reaction-force probe creation, results created before a solve, modal/static checks, mesh size, body sizing, nonlinear settings, contact type changes, material and suppression changes, point-mass conversion (it reproduces the previous implementation's centre of gravity and attachment faces, which shows agreement, not that the inertia is right - see the next point), reports (TXT), non-ASCII names including an emoji, list_point_masses in StandardMKS / CGS / NMM / BIN / UMKS, named-selection creation and dependents, the reuse of identical results and contacts, a connection to a silent socket (refused after about 2 s by the protocol probe, the existing connection untouched), a load scoped to a body-type named selection (refused and rolled back; Mechanical leaves such a load without a scope), the analysis-selection checks (explicit analysis, name, expected type) on a model with three analyses, deleting a named selection by object id, the reparse-point link test (System.IO.File.GetAttributes) on a junction, a regular file and a folder, redaction of paths in replies, and recovery after a transport failure (exit code 0 in a subprocess).

• generate_mesh was only run on a model that already had a mesh (it returned the existing counts at once); meshing a fresh model and the failure path were not exercised live.

• Never run live (needs a solve, which these tests forbid): values read from solved results (max/min, reaction force x/y/z/total, modal frequencies), export_nodal_results on real data, the background solve with a real solver, generate_report(fmt="docx") with real python-docx (tested with a fake docx module), set_2d_behavior on a truly 2D model, and convert_*_to_point_mass on a model whose loads are scoped to the converted bodies (covered by unit tests with the fake model only).

• Standard Earth Gravity only supports +-X, +-Y, +-Z (always 1 g). Any other direction is an error.

• Mechanical 2026 R1 exposes no solve cancel, so a running solve cannot be stopped from here. solve_analysis(wait=false) runs the blocking solve on a background thread; other Mechanical tools report "busy" until it ends, and get_solve_status answers Solving without touching Mechanical.

• Materials are created in Workbench Engineering Data, not by this server; assign_material only accepts names that already exist there.

• Numeric limits (the code enforces exactly these): element sizes 0.001 to 100000 mm, proximity multiplier 0.1 to 100, friction 0 to 10, forces and moments at most 1e12 in magnitude, pressure 1e15 Pa, displacements 1e9 mm, rotations 36000 deg, samples 2 to 5000.

• Point-mass inertia is only exact when the principal axes are the global axes. Mechanical gives a body only its three PRINCIPAL moments of inertia (MomentOfInertiaIp1-3; no tensor, no axes, no products of inertia). They are used as Ixx, Iyy, Izz and the response says inertia_basis: "principal" with warnings: always one, one more when several bodies are combined (their products of inertia are dropped because a point mass cannot take them), and one when planar faces are not aligned with the global axes. A body rotated against the global axes gets wrong Ixx/Iyy; check or set the inertia by hand in that case. Mass, centre of gravity and the attachment faces are unaffected.

• A box selects by centroid. box_mm picks the faces/edges/vertices whose CENTROID lies inside the box, not the entities that lie entirely inside; the response counts the selected entities that stick out in warnings, and fully_inside=true requires the whole entity (face and edge bounding boxes) to be inside.

• Names are not unique in Mechanical (two contact regions can both be "Contact Region"). set_contact_type(region_names=...) matches every region of that name and says so; list_contact_regions shows index and duplicate_names. delete_named_selection deletes every selection of that name.

• Idempotency. The result tools and add_contact return an identical existing object (already_exists: true, created: false, and suppressed - a reused object may be suppressed) instead of adding "Equivalent Stress 2"; pass reuse_existing=false to force a duplicate. A contact region counts as the same when it joins exactly the same entities with the same type and (frictional) friction coefficient. If a region joins them but differs, add_contact refuses and lists the differences; update_existing=true changes the type/friction (restored on failure) and reports updated. Behaviour and formulation are not set by the call; they are only reported in existing_properties. Point-mass conversion refuses to run twice for the same EXACT prefix/part (the mass would count twice) unless allow_duplicate=true: the selection it makes is named NearestFace_<start of prefix>_<hash of the whole prefix>, so SYS\\B and SYS\\Backing never collide and are both allowed. Two different prefixes that match the same body are not recognised as duplicates. delete_named_selection refuses while loads, results, contacts, point masses or mesh controls are scoped to the selection (it lists them) unless force=true. Only objects scoped through a Location, GeometryLocation, SourceLocation or TargetLocation are recognised: Force-reaction probes (their scope is a copy of the entities), joints, springs, remote points, coordinate systems and anything else that references a selection some other way are NOT detected.

• Transport failures. If a call fails in transit (timeout, dropped channel) the script is not re-sent and the response says it may or may not have run. The connection object is dropped outside the exception handler and finalized before its channel is closed: doing it inside the handler made the whole server process crash (exit 139) because the failing frames kept the object alive. The handshake with a port that is not Mechanical is bounded (MECHANICAL_MCP_CONNECT_TIMEOUT_S, default 20 s) and does not hold the Mechanical lock; a late connection that arrives after the timeout is dropped the same way.

• Honest status while Mechanical is busy. During a blocking solve check_mechanical_connection answers connected: true, busy: true, get_solve_status answers status: "Busy" and disconnect_from_mechanical answers busy, each within MECHANICAL_MCP_PROBE_WAIT_S (1 s) plus scheduling. The guarantee is precise: these three tools run on their own small thread pool and never queue for the lock, so other waiting calls cannot starve them. Every other Mechanical tool waits for the lock up to MECHANICAL_MCP_LOCK_WAIT_S (60 s), but at most MECHANICAL_MCP_MAX_WAITERS (4) calls wait at a time; the next ones get busy: true immediately. A connection attempt to a port that never answers takes at most MECHANICAL_MCP_CONNECT_TIMEOUT_S and at most two such attempts may be pending.

• Reconnecting. The remembered port is stored with a hash of the project folder (and its name). The identity is read when you connect: if the project in the session is later saved under another name or a different project is opened in the same session, the remembered identity is stale until you connect again. A silent re-attach is only made while the port still serves that project (otherwise the call fails and tells you to connect deliberately), is reported once as reconnected_to in the tool output, and the port is forgotten when it refuses connections or after three bad answers in a row (tested), not after one; a busy or slow session (handshake timeout) is neither a refusal nor counted. A remembered session with no project identity (an unsaved project) is never re-attached silently: call connect_to_mechanical.

• linearize_path merges coincident nodes (interface nodes of touching bodies; their values are averaged) and triangulates only the nodes near the line (a k-d tree), so 216k nodes take seconds. A line that crosses a hole or notch is a gap error carrying gap_diagnostics (largest span and where it is); allow_gaps=true accepts it and reports interpolated_across_gap. The criterion is local and relative: the simplex a sample falls in is a gap only when its longest edge is more than 3 times the median of the simplices around it AND fewer than 20 % of those neighbours are of similar size. Graded, polar and long-element meshes (10:1, line along the long side or across it) and an abrupt step in element size (a fine region next to a coarse one) are therefore not flagged, but a void smaller than about three element sizes is not detected either. A section line shorter than 0.1 % of the model or than one element is refused. A line end that lies outside the triangulation by less than snap_fraction (default 0.25) of an element - on a curved boundary or a hair outside a flat face - is extrapolated from the nearest simplex and listed in endpoint_adjustments; farther out the line is refused. MECHANICAL_MCP_MAX_NODES (default 1,000,000) and MECHANICAL_MCP_LINEARIZE_TIMEOUT_S (default 60) bound the work; a timed-out run is told to stop (it checks a cancel flag between stages) and only one linearization may run at a time; a second call gets busy until the stopped one has ended. Only the single exported scalar is linearized: the file records its orientation; expect_component makes a mismatch an error (and is reported as component_unverified when the file has none); a line running along another axis than the component is refused unless accept_direction_mismatch=true; components are never combined.

• Export step and time. export_nodal_results exports the result at the step it is currently set to; an analysis with several load steps needs an explicit load_step, and the file records the step, substep and time (step). Time-history results are exported one step per file.

• Exported units. PlotData does not declare its unit. The export compares its largest value with the result's own Maximum and records value_unit_verified; when it is false the unit label is unconfirmed. This has only been unit-tested (a solve is needed to run it for real).

• Explicit Dynamics analyses keep displacement as a time table; the value you set is the last entry (the first is 0 at t=0). set_nonlinear_settings on an analysis without those settings returns a clear error.

• get_solve_status messages come from Mechanical's global Messages pane (all analyses); the response says so.

• One Mechanical session per server process; sessions on other machines are not supported.

• Generated scripts run in a private copy of Mechanical's globals, so nothing leaks between tools; run_mechanical_script (when enabled) shares Mechanical's persistent globals as before.

ACP server (ansys-acp-mcp)

A second MCP server, acp_mcp_server.py, lets an AI assistant build ANSYS Composite PrepPost (ACP) layups and solid models through PyACP (ansys-acp-core), then convert the exported solid mesh into a clean per-fibre-angle external model for a Mechanical coupled-field analysis (for example thermal-electric lightning studies). Every tool returns JSON with ok and a short, specific error on failure.

Tested with ACP 2026 R1, ansys-acp-core 0.3.2, mcp 1.30, Windows, Python 3.10.

Install

pip install -e ".[acp]"

ACP must be installed (the server uses acp_grpcserver.exe from %AWP_ROOT261%\ACP). Set AWP_ROOT261 (or the equivalent for your release) if ANSYS is not in the default location.

{
  "mcpServers": {
    "ansys-acp": {
      "command": "ansys-acp-mcp",
      "env": { "AWP_ROOT261": "C:\\Program Files\\ANSYS Inc\\v261" }
    }
  }
}

Tools

Typical flow: acp_launch → acp_import_model (shell mesh) → acp_create_material → acp_create_fabric → acp_create_rosette → acp_create_oriented_selection_set → acp_create_layup → acp_create_solid_model → acp_update → acp_export_solid_model → acp_prepare_external_model. See examples/build_plate_layup.py.

Notes

• The laminate is extruded toward orientation_point; a shell at z = 0 with orientation_point = (0, 0, -1) puts the plies below the surface.

• acp_create_layup names plies ply03_-45 etc.; acp_prepare_external_model reads the fibre angle from that name.

• ACP solid CDBs are structural (layered SOLID185). The converter gives each angle its own material and coordinate system so a coupled-field solve can use them, and runs the result through MAPDL so Workbench's External Model reader accepts it.

• Hexahedral solid models only; composite shell layups, sub-laminates, draping and taper are not wrapped yet.

Development

python run_acp_unit_tests.py        # CDB / Workbench-script tests, no ANSYS needed
python examples/stdio_smoke_test.py # start the server over stdio and list tools

Integration tests need ACP: set ACP_MCP_INTEGRATION=1 and ACP_MCP_SHELL_CDB to a shell CDB, then run pytest tests/test_integration.py.

License

Apache License 2.0 — see LICENSE