Native, not bolted on
There are two ways to put AI into a coordination tool. The first bolts a chatbot onto it: you ask a question about a clash and get a paragraph back. It can sound clever, but it cannot act on your model, it cannot check what it said, and the next answer may contradict the last.
The second puts the AI where the decisions are, and puts deterministic code on both sides of it. Below the AI, code reads the model and turns coordination knowledge into data: system classes, the project's profile, free space, both runs and every feasible fix with its score. Above it, code builds the chosen fix, Revit verifies it, and a person confirms it. The AI does the part only it can do: weigh the evidence, choose and explain.
That arrangement is what makes the AI safe to use at scale. It never writes geometry, it never sees an option the hard limits removed, and nothing it proposes reaches your model without passing Revit's checks and your click.
1 · Import the report
Open the ClashFree pane from the ribbon and import one or more Navisworks XML clash reports or BCF files; several merge into one list. ClashFree matches each clash to its two elements in your model and its loaded links, merges duplicates across tests, sets aside the clashes that need no action and groups the rest by the run that causes them. Report details at the top of the list shows which file is which model and, for a re-import, what is new and what is gone. (The whole sorting step is in One fix, many clashes.)
No report? Pick the two clashing elements in Revit, Tab to the beam in the link, and the pane opens that clash.
2 · Choose your trade
After the first import, ClashFree asks once which trade you are: HVAC, Plumbing, Electrical, Fire protection, whatever trades the model has, each with how many groups it can fix. It remembers the answer; the trade pill at the top of the list changes it. From then on the list is about what your trade may move:
- To fix: your runs that ClashFree can move.
- Coordination: nothing ClashFree may move; another trade or a person decides.
- Ignored: duplicates, approved or resolved in the report, within tolerance, or set aside by you.
- Cleared: clashes that no longer clash in the model.
3 · Pick the clashes
Tick the groups or clashes to work on. Shift-click selects a range, Space toggles the focused row, Esc clears. Search, the Level dropdown and filters (priority, strategy, AI status, change, related) narrow the list, and the list icon switches between groups and a flat list of clashes. Clicking a row shows it in Revit; its › opens it.
4 · Send to AI
Send to AI adds the selection to the AI queue and shows the cost first: how many items, which model, about how much. Then the queue works ahead, one item at a time, with a slim progress bar and Stop under the header. For each item the scored options are listed, the AI chooses an offset or a short whole-run move, and the fix is applied, verified and rolled back in Revit without changing your model; when a drag, tee move or reroute scores best, the item is flagged with that verb named, for Claude to carry out over MCP. Rows show the outcome as it comes: verified, failed, flagged or stale. Update 2026-10-25: the queue's AI now carries out drags, tee moves, reroutes and area fixes itself, from the same scored options. Send to AI needs no API key; Run does, and Claude over MCP can take queued items instead.
The Queue page shows what is running, what is waiting and the results, and Claude's live budget when Claude is working. A greyed-out action says why on hover, for example when a row is another trade's run.
5 · Keep, reject or revise
Open a verified fix. The clash page shows both elements, the AI's proposal and why, Revit's verification with a snapshot, and the bill of what the fix changes on the run. Preview fix draws the fix's route in Revit as a ghost, with the clash points it clears and any new ones. Keep applies it as one undo step. Reject leaves the model exactly as it was. Or type what you want changed, "go over the beam" or "keep 100 mm from the wall", and get a new verified preview.
Opening a clash re-tests it first. If a fix kept elsewhere already cleared it, the page says so and offers Next; no AI is asked.
Every proposal, from the pane's AI, the queue or Claude, runs inside one Revit transaction group: applied, verified, rolled back, and only re-applied when you press Keep. Reject restores the model exactly. Nothing is committed without your confirmation.
The Cleared tab keeps the list honest
After an import, when the pane opens and after each Keep or Reject, ClashFree re-tests the To fix clashes in Revit in the background: geometry only, no AI, and clashes whose elements did not change are skipped. The header shows the check's progress and the list updates as clashes clear, whether your fix cleared them, another kept fix did, or another trade moved first.
The same loop, with Claude
ClashFree ships an MCP server, and the installer can connect it to Claude Desktop or Claude Code. Claude then works the same loop with the same tools: it lists the groups, opens a clash, reads both runs, the free space and the scored options, and proposes fixes through the same validator and the same verification, with the full set of verbs: drags, tee slides and tilts, shift and stretch, reroutes and area fixes. Each clash has a budget of tool calls and minutes, so a hard case ends in a flag rather than a loop. The pane shows what Claude is doing, with Stop; Keep stays your click.
What the AI actually sees
An AI is only as good as what it is given, and a clash on its own is very little: two element ids and a point. Input is cheap compared with a wrong decision, so ClashFree gives the AI the context a coordinator would look at:
- both runs through their fittings, with every other clash on them and where B runs beside A;
- the free space around the clash, the obstacles and the walls, with their structural use and fire rating;
- the clash test's criterion: a hard clash, or a clearance test and the clearance every fix must keep;
- the coordination profile (ceiling lines, height zones, classes) with the source of each value;
- the system classes of both runs and the cheaper-change reason;
- precedents, twin rooms and the corridor's access zones;
- the scored options, the removed ones with their reasons, and each option's bill of changes.
The AI is told what to cite and what not to claim. Its answer is short: an option id, or a strategy with millimetres, degrees and a direction, plus an explanation. Everything else is code.
What stays human
The decision to keep a fix. The engineering re-checks a fix triggers (airflow, hydraulics, falls, firestops), which ClashFree lists and never claims are fine. The calls that belong to another trade. And the flags, which come with their reasons and a limit code when no option passed. That is not a limitation of the AI; it is the design. See the limits it cannot break.
Ready to try the loop on your own report? The free beta takes a few minutes to set up, and the help center walks through every screen.