toqan

Data work,Done right.

Process data. Build evals. Judge answers. Train models.

Try an agent
1 / 4
01

Playground

Connecting…
API reference ↗

Inspect your files, resolve inconsistencies, and export clean data with reproducible scripts and a quality report.

1. Bring your files. CSV, JSON, JSONL, TXT and more, up to 20 files and 10 MB in total. Drop them here, choose files, or start from a small synthetic sample.

    Advanced options

    Optional configuration passed to the agent. Keep credentials out of instructions and options.

    Held in this tab’s memory only. Never saved to your browser.

    A fresh workspace for each job. Results come back as a ZIP.

    Returns a cleaned dataset, reproducible processing scripts, and a quality report with an artifact manifest. Deliverables depend on your instructions and a successful run.

    Already have a job ID?

    Use the same API key that submitted the job. Closing this page never cancels a job.

    02

    Developers

    Four calls.
    Files to finished.

    Use your API keyAsync · ZIP results

    Send your key as Authorization: Bearer $TOQAN_API_KEY. Keys are issued by the team; there is no self-serve sign-up. Submit returns 202 with a job ID. Poll its status while the agent works. Retrieve returns a ZIP once the job succeeds; before that it returns 409.

    Inference API
    submit.shAGENTS API
    Explore the API reference
    03

    Questions

    Good questions.
    Clear answers.

    How is this different from the inference APIs?

    The inference APIs make predictions on prepared input. Agents do the work around them: prepare the data, build the tests, check the answers, and tune a small model. Use their results in your existing inference workflow.

    What happens when I submit a job?

    A dedicated VM is created from the agent’s prepared template. Your files arrive in that workspace, and Claude Code runs your instruction with the agent’s CLAUDE.md and its subagents. The results are validated, packed into a ZIP, and stored by the API. Then the VM is destroyed. Cancellation, errors, and timeouts also clean up the workspace.

    What is in the results ZIP?

    The agent’s deliverables, an artifact manifest, and a readable report. Depending on the agent, that can be cleaned datasets and scripts, evaluation cases, judgment scores, or a model adapter.

    What does model training support?

    The first setup targets a small Qwen 0.8B model with LoRA, bounded steps, and a separate evaluation split. The template determines the exact model and compute. Small runs validate the setup; they do not demonstrate production model quality.

    What happens if I close this page?

    The API tracks accepted jobs independently of this tab. Closing it does not cancel a job. Keep the job ID to check status later with the same API key.