Evaluations and upgrades
Connecting MCP is available on all plans. Running evaluations uses the same account allowance whether started in the app or through MCP. The sidebar shows your remaining evaluations. The Free plan includes one AI system bench per month from AI Systems. Clicking Start benching starts immediately and includes every detected prompt in that system, including prompts across its repositories. Each prompt retains the Free plan’s 100-case ceiling. This tests the detected prompts and saved context, not full live agent execution or tool side effects. System starts usePOST /api/evaluation-runs/system with {"ai_system_id": 123}. The server resolves the component list, checks access to every repository and shares one Free allowance across the group. Paid plans retain per-prompt metering. The legacy prompt-start endpoint still permits only one prompt per Free request.
On Free, one account credit is reserved for the system bench. It is consumed only when every prompt in the group completes with saved scores. A failed, canceled or partial group releases that reservation. Failed provider calls cannot count as a successful evaluation. A fully cached group does not consume another credit. API-key execution caps still apply; failed attempts do not refund a key’s cap.
Choose Get more evaluations to open secure billing directly. Free or canceled accounts receive checkout for the selected next plan; existing subscribers receive a subscription-update confirmation. Review the price, taxes and proration before confirming payment. View all plans opens the full comparison. A quota error alone does not navigate to billing until you choose Upgrade.
Model-provider selection for improvement suggestions requires active Growth or Enterprise access. It does not change the original model under test or evaluation judges. Existing saved provider restrictions are retained after a downgrade.
Data and retention
Only connect data you are authorized to share. The connection flow links to the applicable processing information. Connecting is an instruction to analyze selected data, not a claim that every legal agreement has been executed. Imported connector evidence defaults to 30-day retention. That limit does not currently describe every document, dataset, evaluation artifact, backup or third-party copy. Remove context through Manage sources and review the applicable DPA and privacy terms for the full processing scope. The expanded deletion path removes an AI system’s structure, saved context and attached datasets and erases its detailed evaluation artifacts before retaining aggregate run history. Raw repository traces need separate deletion or expiry. Source repositories are retained. Manage GitHub access opens GitHub’s installation settings; removing systems inside Bench does not revoke the GitHub App itself. See retention and deletion for deployment status and exact boundaries.Technical controls and a website page alone do not establish GDPR compliance. Organizational responsibilities, executed agreements, subprocessors, transfer safeguards, retention operations and data-subject procedures must also be verified.