Oqoqo

Oqoqo is an evaluation platform for teams whose product is used by AI agents rather than only by people. Instead of scoring a model on a public leaderboard, you define the tasks that matter to your business, run real coding agents against them in isolated cloud machines, and read back what actually happened.

The model it uses

An experiment is a cross product of three things:

• Tasks — the work itself, plus a rubric, the files, and the machine image each run starts from. Files can be uploaded or imported from GitHub, so every trial begins from an identical state.
• Agents — Claude Code, Codex, Cursor and GitHub Copilot are supported today.
• Treatments — the conditions you are comparing. A treatment carries the skills, MCP servers, CLIs and SDKs attached to a run, which is how you measure whether shipping a skill or an MCP server actually made agents better at using your product.

Every run executes in its own reproducible sandbox and captures the full trajectory: tool calls, shell commands, and the point at which the agent gave up.

Reading the results

Results land in a Matrix view for pass rate and lift across the grid, a per-run trace and output view, and Frictions — an automated pass that flags where agents stalled, retried, or stopped. Rubric scoring ("Evals") and friction detection ("Insights") can be switched on at launch or re-run later against completed trials. An assistant called Ask Oqo answers questions about an experiment and can take routine actions inside the project.

Getting it into a workflow

Beyond the web app, Oqoqo ships a CLI and an MCP server, so an agent session can create tasks, launch experiments and read results without leaving the terminal.

Pricing

A free plan grants a monthly allowance of runs per organisation. Paid subscriptions add unlimited team members and projects with a larger monthly run quota, and top-up run packs that do not expire.

Oqoqo launched publicly on 10 August 2026 and was the day's number-one product on Product Hunt.