Adam CAD Harness Integrates with Fusion and Onshape for Agentic CAD Editing

Adam (adam.new) has released a harness that natively integrates with Autodesk Fusion and PTC Onshape, allowing an AI agent to read and edit your existing CAD feature tree. The tool focuses on giving mechanical engineers full visibility and control, rather than being a black-box text-to-STL generator.
Key Features
- Model-agnostic: picks the best frontier model (e.g., GPT 5.5, Opus 4.7) from its internal benchmark per task.
- Lives inside Fusion and Onshape; reads the feature tree and edits it agentically.
- Common use cases: merge redundant features, rename features for readability, round all internal edges with a 2mm fillet, parametrize models, or generate CAD end-to-end.
- Leverages FeatureScript (Onshape) and Python (Fusion) under the hood.
Installation on Autodesk Fusion
One-liner install (recommended, ~10 seconds):
macOS: paste in Terminal
curl -sSL https://fusion.adam.new/install.sh | bash
Windows: paste in PowerShell
irm https://fusion.adam.new/install.ps1 | iex
After install: quit Fusion, reopen, press Shift+S → Add-Ins tab → find AdamFusion → click Run and tick 'Run on Startup'.
Manual install: download AdamFusion.bundle.zip, unzip, move the Contents folder into:
- macOS:
~/Library/Application Support/Autodesk/Autodesk Fusion 360/API/AddIns/(rename to AdamFusion) - Windows:
%APPDATA%\Autodesk\Autodesk Fusion 360\API\AddIns\AdamFusion
Then quit and reopen Fusion, enable via Add-Ins panel, and sign in with your Autodesk account.
Context
Adam’s earlier text-to-CAD work is open-sourced on GitHub. The team also notes that the Anthropic Autodesk connector validates the space but differs in being model-specific; Adam remains model-agnostic.
📖 Read the full source: HN AI Agents
👀 See Also

SkillOpt: Optimizing Markdown Skill Files as Trainable Parameters for AI Agents
SkillOpt formalizes the ad-hoc process of editing markdown skill files for AI coding agents, using frontier models to propose bounded edits gated against validation sets. Best skills converge with 1-4 accepted edits out of many proposals, and transfer across models like Codex to Claude Code.

Claw Code Agent: Python Reimplementation of Claude Code Architecture for Local Models
Claw Code Agent is a Python reimplementation of the Claude Code agent architecture that runs with local open-source models through OpenAI-compatible backends like vLLM and Ollama, featuring tool calling, slash commands, and tiered permissions.
Needle: A 26M Parameter Tool-Calling Model Built Entirely Without FFNs
Needle is a 26M parameter function-calling model with no MLPs, achieving 6000 tok/s prefill and 1200 tok/s decode on consumer devices. It beats FunctionGemma-270M, Qwen-0.6B, Granite-350M, and LFM2.5-350M on single-shot tool calling.

Tilde.run: An Agent Sandbox with a Transactional, Versioned Filesystem
Tilde.run provides isolated, reversible sandboxes for AI agents, with a versioned filesystem that mounts GitHub, S3, and Google Drive, and network isolation by default.