Driving AEC software with a coding agent
A coding agent (Claude Code, Cursor, Copilot Workspace) reads your instructions, opens tools, runs scripts, and patches files -- without you clicking through menus. For a sole-practitioner A/E firm, that means automating the tedious parts of production: exporting schedules, querying a model, batch-converting drawings, or pulling cost data directly into a report. Whether a tool can be driven this way -- and how well -- is what Agent Friendliness (C2) measures in our rubric.
Bottom line up front: The access surface a tool exposes determines what a coding agent can actually do with it. An MCP server or CLI means you (or an agent you point at the problem) can automate the tedious parts in minutes. A GUI-only tool means every interaction still needs a human at the mouse.
Five questions before you trust AI output on a project
Many tools market "AI" heavily. These five questions cut through it.
A tool's own "AI" claim is marketing copy -- the aiFeatures field in our catalog
is a factual description of what the model actually does (and who provides it), not a score.
A summarization widget over your own data is not the same as a trained domain model.
Know the task the model performs (clash detection, spec lookup, quantity takeoff) and whether the vendor discloses which foundation model it uses. Undisclosed models make independent verification impossible.
AI errors are your professional liability, not the vendor's. Verify every output against the source before it goes into a set. The tools that make verification easiest -- by showing sources, citing sections, or producing an auditable output -- are the ones worth trusting in production.
A solo practice has no second reviewer. Prefer tools where AI suggests and you confirm, over tools that write or change files autonomously. In a one-person shop an unchecked automation that deletes or overwrites is a single point of failure.
Project data fed into a vendor's AI may be used to train future models or stored offshore. This is Red Flag 6 (RF6) in our catalog. Check the vendor's data policy before uploading drawings, specs, or client information to an AI feature. Many vendors now offer opt-outs -- but you have to ask.
The access ladder: why it matters for a one-person firm
Every tool sits somewhere on this ladder. Higher rungs mean a coding agent -- or a simple script you write once -- can automate repetitive work that would otherwise take hours at the GUI. Lower rungs mean a human must still click every step.
Agent-drivable tools in the catalog
42 tools with C2 ≥ 4 or an MCP server / CLI, grouped by job-to-be-done. C2 is the Agent Friendliness score (1–5). Badges show the highest access surface available.
Win & Set Up the Work
Design the Building
Engineer & Analyze
Produce & Issue Documents
Coordinate & Collaborate
Administer Construction
Run the Practice
C2 measures how easily a coding agent can drive the tool. Scores of 4–5 mean a solo practitioner can realistically automate repetitive tasks without vendor help. See how we score for the full rubric.