Recommended path · about 30 seconds

Hand the benchmark to your AI agent.

No output copy-pasting and no long forms. Give your agent one instruction; it fetches Mahak prompts, preserves responses verbatim, and submits reviewable run metadata.

Start: 30sec
Instruction: 1
Manual paste: 0

01 / 03

Start with the one-instruction path

Choose a model Mahak still needs

A · FASTEST

Agent skill

Your agent reads the published contract, selects prompts, preserves outputs unchanged, and submits through the API.

Paste this into your agent

Read https://mahak.waqf.dev/skills/mahak-contribute/SKILL.md and benchmark <model-slug> on Mahak starting with prompt p_16

Read the skill contractClaude Code · Cursor · Codex · ChatGPT · OpenCode

B · CLI

A recoverable terminal run

The CLI creates a prompt-version-bound manifest, saves progress, and submits ready files in batches. Your agent generates the responses; the tool never impersonates a model.

curl -fsSL https://mahak.waqf.dev/install | sh
mahak start --model <model-slug>
# Your agent writes the responses, then:
mahak sync

The installer currently supports Linux x86_64 and installs into ~/.local/bin.

C · MCP / API

Direct provider integration

Connect an evaluation platform or CI pipeline through native MCP. Personal access tokens travel in the Authorization header and never enter model context.

Endpoint

https://mahak.waqf.dev/mcp

Mint a personal access token

THE ONE MANUAL CONTRIBUTION

What should Mahak test next?

Request a new domain or category, or propose a realistic Arabic scenario for an existing domain. Mahak reviews every suggestion before publication.

Suggest a domain or scenario