Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.cekura.ai/llms.txt

Use this file to discover all available pages before exploring further.

Use OpenAI’s Codex CLI together with the Cekura MCP to design evaluators, run them against your voice agents, and triage the results — all from your terminal.

​
Quick Guide

1

Connect Codex to Cekura

Add the Cekura MCP server to Codex. Edit ~/.codex/config.toml and add:
[mcp_servers.cekura]
url = "https://api.cekura.ai/mcp"
On first use Codex opens a browser to authorize against your Cekura dashboard. For API-key-based auth (project-scoped credentials, shared CI access), use the CLI form instead:
codex mcp add cekura --env CEKURA_API_KEY=YOUR_API_KEY_HERE -- \
  sh -c 'npx -y mcp-remote https://api.cekura.ai/mcp --header "X-CEKURA-API-KEY:$CEKURA_API_KEY"'
Verify by starting Codex and asking: “List my Cekura agents.”
2

Set up mock data and cleanup hooks

Complex agents depend on database state — users, accounts, sessions, entitlements. Stand up a lightweight webhook server that seeds mock data before a run and listens for Cekura’s post-run webhook to reset the database. This keeps concurrent tests isolated and repeatable.
3

Let Codex learn the schema

Run Codex from your agent’s repo so it can read your tool definitions, prompts, and database schema directly. Then prompt it to generate scenario coverage through the Cekura MCP — for example: “Read the agent’s tools and prompts in this repo, then create 10 Cekura evaluators covering the most common user flows.”
4

Run the full suite

Kick off all evaluators at once through the MCP: “Run every evaluator on agent X and report pass/fail.” Cekura executes them concurrently and surfaces pass/fail with full transcripts and traces.
5

Triage with Codex

Ask Codex to pull failing runs and classify them: “Fetch the failures from the last run and tell me which are real bugs vs. flakes.” Because runs are non-deterministic, rerun each failing test 3–4 times and ask Codex to compute a pass rate. One pass out of four usually points to a prompt or tool-routing bug, not infra.
6

Fix in-place and re-run

With both your codebase and the Cekura MCP in the same Codex session, you can go from failing transcript → suspected prompt or tool bug → edit → re-run the failing evaluator without leaving the terminal.
A good first prompt: “Use the Cekura MCP to list my agents, then read the prompt for the first one and propose 10 workflow evaluators covering the most common user flows.”