Cosmon

Nexus writes into named engineering applications. docs.cosmon.com says: "Connect Nexus to SolidWorks, Ansys, COMSOL, Abaqus, or Star-CCM+." The SDK adds Fluent and Ansys Mechanical, the change log adds Fusion 360 and Catia V5, the marketing pages add NX, and Jama and Cameo are named on the requirements side: eleven commercial applications across CAD, CAE solvers and requirements or MBSE tooling.

The writes are consequential: automated geometry edits, parametric variations, dimensioned drawings with GD&T and BOMs authored natively inside SolidWorks, mesh generation, material and boundary condition assignment, simulation runs and PDF export. Changing geometry in a customer's CAD model is a write into their system of record, and a riskier one than a CRM update.

The integrations are versioned and maintained. Release 0.1.22 "moves the SolidWorks FEA agent into production, enables Fusion 360 to author drawings and export PDFs" and "expands validated workflows for CATIA V5", alongside corporate proxy support, and a COMSOL AI agent shipped separately in August. No integrations page, supported version matrix or connector catalog is published.

Sourcedocs.cosmon.com front page, batch22a and batch19a change entries, pypi.org cosmon-agent-sdkread 2026-09-13

Nexus orchestrates across tool boundaries, which is the product's whole thesis. Cosmon positions engineering AI as living above the individual platforms and automating "not just tasks within each silo but the transitions between them": requirements tool to CAD to solver to PLM. A published closed loop runs the full chain, from a requirement change in SysML through CAD updates and simulation reruns under governance to a traceable review package.

The steps are documented individually: turning natural language requirements into exact boundary conditions and pass or fail criteria; physics aware geometry cleanup for meshing that keeps critical structural features; material and boundary condition assignment; solver runs; results interpretation; and report drafting, each with its own docs section under CAD or CAE.

The SDK shows them composing: a single client.run() call such as "Run a static study on bracket.sldprt" emits a sequence of typed ToolCall events, and runs "can work for many minutes". Agents are addressed by name (agent="solidworks"), so work is routed to specialized actors, and a SolidWorks FEA agent and a COMSOL agent moved into production separately.

Because the SDK lets a buyer batch, script and embed Nexus in their own pipelines, the customer composes the sequence and its surrounding logic in their own Python rather than accepting a fixed pipeline.

Sourcedocs.cosmon.com CAD and CAE sections, pypi.org cosmon-agent-sdk run model, cosmon.com Engineering OS positioningread 2026-09-13

Nexus indexes engineering artifacts, not just documents. Cosmon describes indexing "the actual physics and geometry of files" so engineers can query prior work conceptually, for example past thermal analyses of battery packs with a specific cooling geometry: retrieval keyed on geometry and physics against a stored corpus.

The docs open with "Nexus is an AI agent that works directly inside your engineering software. It understands your designs, runs your simulations", Cosmon names "multimodal geometry technology that understands CAD structure, physics, and design intent", and the SDK shows it working: "What's the max von Mises in bracket.sldprt?" is answered from a model file.

The grounding sources are the customer's systems of record: geometry in CAD files, results in solver output and requirements in Jama or Cameo, with PLM metadata interpreted so certification simulations run only on the correct released revision.

Natural language requirements become exact boundary conditions and pass or fail criteria, so the numbers come from the requirement document rather than the model. Outputs are "grounded in the customer's engineering sources of truth" and produce "a traceable review package", so a result can be traced back to the requirement and revision it rests on. No index technology, embedding approach or retrieval mechanism is disclosed.

Sourcedocs.cosmon.com front page, cosmon.com multimodal geometry section, pypi.org cosmon-agent-sdk worked exampleread 2026-09-13

Oversight is a documented setting, and the interactive path is review and approve. In the SDK, "attended vs unattended: interactivity is set when you build the client, by whether you supply handlers", so a caller decides in code whether the agent may proceed alone or must call back to a person.

Cosmon states the gate per workflow: "Nexus diagnoses solver issues and proposes fixes; you review and approve before any change is made"; "Nexus flags manufacturability, assembly risks, and cost issues; you review findings and decide"; "Nexus interprets simulation results and drafts reports, ready for your review and sign-off." Engineers grant permission for actions, and even the closed loop case, where a requirement change propagates through CAD updates and reruns simulations under strict governance, runs "without human intervention until final sign off". Every AI decision is also reversible. Cosmon's framing: "Shift from execution to direction. You decide the direction. Nexus executes."

There is no approval queue, no risk tiering by action, no record of who approved what, and no organization level policy forcing attended mode: whoever writes the client chooses, so an engineer can opt out of their own oversight.

Sourcepypi.org cosmon-agent-sdk attended versus unattended, cosmon webinar workflow descriptions, docs.cosmon.comread 2026-09-13

No attestation is published, but a thin per user access model is documented. There is no SOC 2, ISO 27001 or GDPR statement, no named auditor or report and no trust center. Local deployment keeping IP inside the network is a deployment property, interpreting PLM revision rules is integration with the customer's PLM, and keeping designs confidential is a promise rather than a mechanism.

The SDK documents an identity model that is a real, customer facing control: "your login, your credits", and "your code never holds a key", with capability bound to the signed in installed app rather than a shared secret, and Developer access as a setting a user must turn on before the local MCP server can be driven. No SSO, SAML, RBAC, SCIM or admin console is documented.

Sourcedocs.cosmon.com, pypi.org cosmon-agent-sdk authentication modelread 2026-09-13

Each run exposes a step by step trace, and agent actions can be undone. Cosmon states that "every AI decision stays visible, traceable, and fully reversible", and the SDK supplies the mechanism: iterating a Run "yields typed ToolCall / TextChunk events", with the documented example printing each tool step as it happens (print(f"[{event.summary}]")) alongside the answer streaming token by token. A caller can watch which tools the agent used, in order, with a summary of each, while the run is in progress.

Reversibility means an agent action on a CAD model can be undone, which requires the platform to know what it changed; on a product that edits geometry in a customer's system of record, that is the auditability that matters. A closed loop change produces a traceable review package that keeps the record of what was done for sign off. Coverage reports, solver diagnostics and manufacturability findings describe the customer's part, not the agent.

No retention period, persisted run history a customer can browse later, or admin audit log across users is documented; the trace is available to whoever calls a run.

Sourcepypi.org cosmon-agent-sdk event streaming, cosmon.com traceability and reversibility claimsread 2026-09-13

Indexing and recalling past analyses as institutional knowledge is a retrieval corpus about the customer's designs, which is the knowledge layer, not state the agent keeps about its own work. Nothing describes an agent carrying state across runs: no memory scope per user, project or agent, no published lifetime or retention, no expiry or purge path, and no way to read or write accumulated state. Nothing documents recalling a previous run as distinct from a previous file.

The SDK points the same way. A run is created by Client.run(), yields typed events and returns an answer; nothing in the documented client persists, names, resumes or references an earlier run, and attended or unattended mode is set per client rather than carried forward.

Sourcedocs.cosmon.com, pypi.org cosmon-agent-sdk run modelread 2026-09-13

The customer chooses where Nexus runs: "Nexus can run in the cloud or locally so proprietary engineering IP never leaves the customer's network." The product shape backs it up. Nexus ships as an installed desktop application: the release notes live at a desktop app subdomain, the SDK needs an "installed, signed-in Cosmon app", and the capability comes from "the CAD software on your machine".

Software running on the engineer's own workstation against locally installed CAD is the software being where the buyer is, not a residency promise about a vendor cloud. Release 0.1.22 "adds corporate proxy support for managed networks", which a vendor builds only because customers run the product inside controlled corporate networks. Pre release CAD geometry and simulation results are among a hardware company's most closely held assets.

No region, region list, country or residency commitment is named for the cloud option, and no installation or architecture documentation describes what local deployment involves.

Sourcedocs.cosmon.com, pypi.org cosmon-agent-sdk installed-app model, batch22a change entry on corporate proxy supportread 2026-09-13

Named per tool agents can be selected, but no catalog exists. Agents are addressed by name in the SDK (client.run("...", agent="solidworks")), so a buyer chooses which agent does the work, and the change log names them as separate shipping units: "the SolidWorks FEA agent moved into production", and a COMSOL AI agent released in August.

Remove the COMSOL agent and the SolidWorks FEA agent is still a whole product doing unrelated work. The docs group capabilities into families: CAD (modeling, drawings, DFMA, tolerance stack up and reverse engineering) and CAE (simulation preprocessing, troubleshooting and postprocessing).

There is no agent gallery, page per agent with a lasting URL, template library, enable or install action, or entitlement language; documentation sections describing what the product can do are not a set a buyer picks from. Custom build outs are professional services.

Sourcedocs.cosmon.com CAD and CAE sections, pypi.org cosmon-agent-sdk agent parameter, batch22a and batch19a change entriesread 2026-09-13

Nexus can run unattended from a customer's pipeline, but whether it starts on its own from an engineering event is not confirmed. On its insights pages, Cosmon describes event driven change propagation within the toolchain: "a requirement change in SysML can autonomously trigger CAD updates, re-run simulations under strict governance, and generate a traceable review package".

The docs describe two ways in, both started by the caller: "Work with it interactively in the app, or drive the same agent from your own Python with the Agent SDK." Unattended runs are documented, and the SDK exists to "batch, script, and embed it in your pipelines", so a pipeline can invoke Nexus on a commit or a nightly build with no person involved, though the scheduling belongs to the customer's CI rather than to Cosmon. The docs show no mechanism by which Cosmon itself watches a requirements tool or PLM revision and fires.

Sourcedocs.cosmon.com invocation paths, pypi.org cosmon-agent-sdk unattended moderead 2026-09-13

The customer does not choose the model, and none is named. Cosmon says "Nexus goes beyond the average AI experience with multimodal geometry technology that understands CAD structure, physics, and design intent" and that it is "trained on thousands of cross-domain synthetic engineering problems", which describes proprietary geometry models. A vendor that trains its own models has a coherent reason to fix the stack.

No model selector, admin entitlement, per agent or per run model setting, disclosed routing, or bring your own model or key path is documented. The SDK's run signature takes a prompt and an agent name and no model argument. Not one provider, model family or version is named; the site contrasts Nexus with "generic LLM-based agents" without saying what sits underneath. A hardware buyer sending pre release geometry cannot learn which model reads it.

Sourcecosmon.com multimodal geometry section, docs.cosmon.com, pypi.org cosmon-agent-sdk run signatureread 2026-09-13

Cosmon publishes a Python SDK for driving its agent. docs.cosmon.com carries Agent SDK as one of five top level cards ("Drive Nexus from Python: batch, script, and embed it in your pipelines"), with its own documentation section at /sdk/overview, and the front page names the two ways to use the product: "Work with it interactively in the app, or drive the same agent from your own Python with the Agent SDK."

The package is first party: cosmon-agent-sdk on PyPI, published from support@cosmon.com, four releases through 0.1.3 in July 2026, installed with pip install cosmon-agent-sdk. It documents a client surface: Client(), client.run(prompt, agent=...), a Run object with.text(), typed ToolCall and TextChunk events for streaming, attended or unattended mode set by whether handlers are supplied, and a cosmon doctor diagnostic command.

The architecture points inward: "Under the hood the SDK is a small MCP client: it discovers the running app, connects to the local MCP server it exposes, and drives the agent through one run tool." The installed application exposes a local MCP server that the customer's own code drives, once the user turns on Settings > Developer access. The credential model is sound: "Your code brings the caller; the installed app brings the capability (your login, your credits, the CAD software on your machine). Your code never holds a key." The SDK is marked alpha at 0.x, with breaking changes expected.

Sourcedocs.cosmon.com Agent SDK section, pypi.org cosmon-agent-sdk published by support@cosmon.comread 2026-09-13

Nothing tests the agent itself. Cosmon's testing vocabulary is about the customer's design: running a simulation to prove a bracket survives a load, flagging manufacturability and draft issues, checking tolerance stack up and diagnosing a solver that failed to converge all evaluate the customer's part, not Nexus. Design validation, issue detection and holistic optimization are not agent evaluation.

No harness for agent output, scored cases, regression suite over generated geometry, golden set, judge, comparison between agent configurations or accuracy figure for the agent's own work is published. What is published is a performance claim without a measurement: "trained on thousands of cross-domain synthetic engineering problems, providing best-in-class outcomes on real engineering work when generic LLM-based agents fall short", with no benchmark, comparator or number. Being able to undo a decision is control after the fact, not a test before it.

Sourcecosmon.com product claims, docs.cosmon.comread 2026-09-13

Nexus drives desktop CAD on the engineer's machine, but nothing says it does so through the screen. It ships as a desktop application and "works directly inside your engineering software", connecting to SolidWorks, Ansys, COMSOL, Abaqus and Star-CCM+. The SDK says the capability is local ("the installed app brings the capability (your login, your credits, the CAD software on your machine)"), and the change log adds a Nexus Connector and corporate proxy support for managed networks. An agent driving locally installed desktop CAD comes close to local computer control.

What decides it is the interface. Every CAD and CAE application named here, including SolidWorks, Ansys, Abaqus, COMSOL, Fluent, Star-CCM+, Fusion 360 and Catia, exposes a scripting or automation API, which is the ordinary way software drives them, and the published outputs are what those APIs produce: "fully dimensioned engineering drawings, multi-views, sections, GD&T, and BOMs natively inside SolidWorks", mesh and boundary condition setup, and PDF export. Nothing first party describes screen capture, cursor control, pixel reading or targeting interface elements.

Sourcedocs.cosmon.com front page, pypi.org cosmon-agent-sdk, cosmon.com SolidWorks drawing outputsread 2026-09-13