Where agents work on 3D data
An agent cannot pull down hundreds of gigabytes to answer one question. On CoDim it sends the question to the data, the work runs in the cloud next to it, and the agent gets back only the answer. Your team sees it in the live 3D scene.
Which of the three bracket designs cuts peak stress the most?
I'll compare the three designs in the Bracket study project.
codim dataset_list(Bracket study)
Baseline, Design A, Design B
codim dataset_describe(von_mises)
Peaks 1007, 927 and 782 MPa in the vertical load case
codim scene_actor_set(von_mises)
Side by side, one color scale, 0 to 350 MPa
codim view_camera_set
All three brackets in view
codim annotation_create(Design B)
Peak pinned at the inner fillet, 782 MPa
Design B cuts peak stress the most: 782 MPa at the inner fillet, 22% below the baseline at half its mass. Each peak is annotated in the scene.
Inspect
Dataset overviews, field statistics, and spatial summaries, returned as typed results sized for a context window.
Analyze
Probe points and lines, locate extremes, compare fields between meshes, and run filters from the catalog in the cloud.
Drive the scene
Create scenes, add datasets, set color maps, opacity, lighting, and camera, and capture the live view as an image.
Report
Post findings to the scene's threads, where the team reviews them next to the 3D data.
The agent sends the question. The data stays put.
Compute runs next to the data
Filters, probes, and comparisons run on cloud compute beside the stored data, at full resolution.
Answers sized for a context window
The agent gets back typed results and statistics, not hundreds of gigabytes of mesh.
The work runs once
Each result is kept, so every agent and person on the team reuses it.
An agent probes the same mesh your engineers do
A probe returns every field at a point, from the cell that contains it, on the full-resolution data. The agent gets the numbers, and your team sees the same point in the scene.
Tested methods, not generated ones
An agent on CoDim calls the same filters and probes your engineers use. It does not write its own analysis code, so a result is not off by a tolerance nobody checked.
The questions agents get asked
- StructuresWhere does stress peak on this bracket?
- AerodynamicsWhere does the flow separate on the wing?
- MiningWhat is the ore grade in this block model?
- ScansWhat changed between these two scans?
Every answer has a record
Each result traces back to the data and the operations that produced it, so a person can check what an agent did before it goes into a report or a filing. An agent sees only what its user can see and changes only what that user can change.
- Agents sign in with OAuth and act as the user, with that user's permissions
- Scopes limit an agent to reading, analysis, scene edits, or collaboration
- Changes to shared scenes wait for the user to approve them
- Every tool call is logged, and scene edits are marked as made by an agent
- Text people typed into the data reaches the agent marked as untrusted
- Responses are capped, so a large dataset cannot flood the context window
Works with the agents your team uses
The CoDim MCP server is a remote server with standard OAuth sign-in. Connect Claude, Cursor, or any client that speaks MCP, and we turn it on for your workspace. Plans are sized by data, with unlimited users and agents.
Point an agent at the dataset that breaks your stack.
Send us a run or a scan your team struggles to open. We will load it into CoDim and connect your agent to it.