Use YOLOZU with an LLM or agent

Use an agent to check prediction files and evaluate them against labeled data. Start with a local example, then connect an MCP client.

First successful evaluation

The CPU tutorial creates sample inputs, validates them, and runs COCOeval. Installation needs network access; the example needs no model or dataset download.

Do not infer model quality from a synthetic proof or a successful dry run. Real COCO metrics require the coco extra and labeled inputs.

Inspect before connecting an MCP client

After the isolated installation in the 30-minute path, install the optional MCP extra in that same environment. These commands exit after writing local discovery and sample files; they do not call a provider API or download a model.

Check the installed CLI and MCP schemas before running commands; the installed package may differ from these docs.

python -m pip install "yolozu[mcp]"
mkdir -p reports
yolozu-mcp --help
yolozu-mcp --print-tools --guaranteed --ids-only > reports/mcp_tool_ids.json
yolozu-mcp --sample-generate-config > reports/ai_generate_config.json
yolozu-mcp --sample-review-config reports/ai_generate_config.json > reports/ai_review_config.json

Inspect the returned IDs and review result. The guaranteed-tool filter narrows this discovery output, not the server's runtime surface or filesystem permissions.

Do not treat registered MCP tools as guaranteed execution. Runtime, dependency, data, and permission requirements still apply.

Connect an authorized local client

Configure the client to launch one of these alternatives with the installed environment's executable and an explicitly chosen workspace. These are long-running stdio servers, not one-shot validation commands. Stop the server when the session ends.

yolozu-mcp --transport stdio
yolozu-mcp --transport stdio --surface image-service

Use the full surface only when the broader tools are intended. The image-service alternative exposes only capabilities, asset upload, submission, status, and cancellation. Ask for capabilities first; never reinterpret abstention as successful inference.

The packaged adaptive image service abstains until CNN license review, quality qualification, and activation are completed separately. It is not a ready-to-run hosted detector.

Two starting prompts

Prompt: Validate my existing detection predictions strictly with YOLOZU; report errors without repairing the input.

Expected: Confirm the user-selected local input, inspect the installed schema, then use strict validation. Do not invent a path or enable repair.

Prompt: Show the YOLOZU image service capabilities without running inference.

Expected: Use image_service_capabilities on the image-service connection. Do not upload an image or submit execution.

Machine-readable references

These docs use source version 4.11.0. Source links can change. This build records its sources in source_sha256 and provenance.json.

Install the checked release, generate a sample, strictly validate it, run COCOeval, and try a portable release gate. Synthetic metrics check the workflow.

Japanese instructions for the same generated-sample evaluation and release gate; no model or external dataset download is needed.

Choose a package extra and distinguish installed commands from repository scripts.

Command inputs, examples, side effects, dependencies, and maturity.

Exact generated tool surfaces and MCP input schemas; registration is not execution qualification.

Machine-readable strict validation results, errors, and repair disclosure.

Typed input, result, and error interfaces for local validation and evaluation.

Client connection options and workspace boundaries.

Capabilities, opaque assets and jobs, limits, and abstention semantics.

Prepare a local plugin from a reviewed checkout. It is not a public ChatGPT directory listing. Provider authentication, hosting, and attachment handoff are separate deployment work.

Keep Stable, Bridge, Benchmark, and Research claims separate.

Publish the bundle, verify access, and measure successful use without inferring adoption from crawler traffic.