Documentation

Connect an AI client with MCP

On this page

Connect an AI client to the workbench through MCP, then iterate on a workflow one change at a time without handing over the camera.

What MCP adds

MCP (Model Context Protocol) lets an AI client call the workbench's supported tools: inspect a workflow, propose a candidate and read back the experiment evidence. It is a second way to drive the same task service. It does not add a camera driver, a new image engine or a production qualification.

Question	Answer in the reviewed preview
Where does the client run?	On the Windows workstation. The client starts the local task service itself and talks to it over standard input/output.
Does it change a camera?	Not by itself. Reading a workflow, running an offline experiment and writing camera settings are separate operations. Camera-side tuning still needs the supported device workflow and its verification steps.
Does it need the internet?	No. Remote MCP access is a separate, authenticated configuration; see Permissions, security and data.

Before you connect

Install and start the Windows workbench as described in Installation and configuration, and run one offline workflow so you know what a normal result looks like.
Decide what may leave the workstation. A client forwards the tool results it receives, including image samples, to its own provider.
Prepare a labelled sample set with the expected answers. Keep development images apart from a frozen validation set.
Plan to connect the demonstration (simulated) service first.

Export the client configuration

In the desktop application, open the tuning window and use its MCP configuration export.
Choose the client format: JSON for Claude or TOML for Codex.
Choose the normal service, or the demonstration (simulated) service.
Copy the text or save it. The suggested file name is deyivision-mcp.

The exported entry is named deyivision. It contains the full path of the installed task service, the --mcp argument (plus --simulation for the demonstration service) and three environment variables: DEYIVISION_TUNING_INSTANCE, DEYIVISION_TUNING_DESKTOP_INSTANCE and DEYIVISION_TUNING_DATA. It never contains a token, a camera credential or any other environment variable. The demonstration entry leaves the data folder out because that service keeps its own.

Install it in the client

Merge the deyivision entry into the client's MCP configuration: the mcpServers object of Claude's JSON file, or Codex's TOML file. Where that file lives depends on the client and its version; follow the client's own documentation.
Restart the client so that it reads the entry.
Ask the client to list the DeYiVision tools. If none appear, check the path and the client's log before changing anything in the workbench.
Export again after you move or reinstall the workbench, change the instance or change the data root. The entry stores absolute paths.

Check the first connection

Start with the demonstration service. A successful tool call proves that the client can launch the service and talk to it. It does not exercise a camera or real image inspection, and the demonstration service keeps its data apart from device-mode data. Do not point it at the device-mode data folder: two services would contend for the same task ledger.

Move to the normal service only after the demonstration call works. A tool list is not evidence of a camera connection; see Camera connection and acquisition.

Iterate one change at a time

Fix the sample set and the expected answers before the first run.
Ask the client to inspect the current workflow revision and its latest result. Record the revision, the node settings, the input identity, the output or decision, the elapsed time and any mismatch.
Change one candidate: one supported node or parameter. Submitting a candidate saves a proposal; it does not run it.
Run the candidate on the same fixed images. A run consumes the task's PC trial budget; it is not a new camera exposure or a camera-parameter write.
Review the evidence yourself: images, output, version and failure cases. A cleaner picture can still give the wrong decision.
Keep or discard the candidate, then repeat. If a person or another client changed the workflow meanwhile, read the latest revision and its human review first. An old draft must not overwrite the current revision, and a candidate tied to an outdated review revision is refused.

See Build and run a workflow for the revision, submit and run controls.

What stays under human control

Camera writes. An MCP client does not bypass the controls in Jobs and inspection results. Writing a camera still needs the right device and session, permissions, an eligible state, a recorded baseline and a readback.
Permissions. Model access is not inherently read-only: an authenticated model can have write-capable tools. Decide what the client may reach before you connect it, and review what it did afterwards.
Unknown results. If a write has an uncertain result, read the original request receipt. Do not repeat the write under a new request ID.
Acceptance. An offline experiment on saved images does not prove how a new frame behaves on your line. Agree samples and acceptance limits before you make an operational claim.

Data and privacy

The client sends what the tools return, including image samples, to the provider you selected. Review the recipient, the data rights and the provider's policy first. A request-level store preference is not a guarantee about every downstream party's retention.

The exported configuration holds local paths and folder names only. Startup output can contain local access tokens, so do not paste it into a support message.

Troubleshooting

Symptom	Checks and next step
The client lists no DeYiVision tools	Confirm the entry is merged into the client's real configuration file, restart the client, and check that the command points to the installed task service executable. Re-export if the workbench moved.
The service starts and then fails	Compare the instance and data variables with the running workbench. Use the matching pair, and never share one data folder between the demonstration and the normal service.
It worked, then stopped after a reinstall or move	The entry stores absolute paths. Export the configuration again and replace the old entry.
A candidate is refused as outdated	Read the latest workflow revision and its human review, then propose the change again.
A write has an unknown result	Read the original request receipt before deciding whether a new operation is appropriate.

Limits of this guide

The exact tool list and tool names come from the connected service and can change between builds; ask the client to list them. The export covers the Claude (JSON) and Codex (TOML) formats; behaviour with each client version has not been qualified here.

A 23 September 2026 engineering check on an isolated Windows build retrieved a source image and a processed image through a remote MCP client. Both runs returned NoRead (0/1). It demonstrated workflow handoff, revision control and evidence retrieval; it did not establish barcode read rate, live camera compatibility or production stability.