Cekura AI Skills

AI-powered skills for building and improving voice agent tests and metrics on the Cekura platform. Works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, and other AI coding assistants.

Table of Contents

• What's Included
• Prerequisites
• Quick Install (npx skills add)
• Claude Code (VS Code)
• Claude Code (Terminal CLI)
• Codex
• Cursor
• Gemini CLI
• GitHub Copilot
• Windsurf / Other Agents
• MCP Server
• Quick Reference
• Platform Compatibility
• Reinstalling Cekura skills
• Links

---

What's Included

14 Skills, 15 Commands in one plugin

Skills	Commands
cekura-coordinator	cekura-onboarding, setup-mcp, upgrade-skills, report-bug
cekura-onboarding	create-metric, list-metrics, evaluate-calls, improve-metric
cekura-create-agent	manual-create-update-eval, autogen-eval, list-evals, run-evals, eval-results, cekura-report, cekura-results-report
cekura-self-improving-agent
cekura-metric-design
cekura-metric-improvement
cekura-predefined-metrics
cekura-eval-design
cekura-personality-design
cekura-infra-test-suite
cekura-bot-test-writer
cekura-flag-call-log-failures
cekura-generate-scenarios
cekura-results-report

cekura-self-improving-agent covers both "improve my agent / auto-tune from eval results" and "fix a production call bug end-to-end". It works on dashboard-managed providers and on custom stacks — config in your own repo, database, or prompt registry, provider agents created at deploy time, custom mock servers — via a per-project capability manifest.

These encode best practices from real client deployments — proactive guardrails, real transcript grounding, iterative improvement loops, coverage planning, and anti-pattern detection.

Stuck after an upgrade? If /upgrade-skills reports stale plugin entries or /plugin install fails with Source path does not exist, do a clean reinstall — see Reinstalling Cekura skills below.

Prerequisites

• Cekura account — Sign up here. Sign in via OAuth or use an API key.
• Cekura API key (only for the Claude Code plugin path / programmatic MCP access) — Found under Settings > API Keys in the Cekura dashboard. Not needed for npx skills add.

---

Quick Install (npx skills add)

The fastest way to get Cekura skills into any Agent Skills–compatible client (Claude Code, Cursor, Codex, Windsurf, OpenCode, and many more).

Install

npx skills add cekura-ai/cekura-skills

The CLI prompts you to pick which skills to install and which agents to install them into. To install everything non-interactively:

npx skills add cekura-ai/cekura-skills --all

Update

# Refresh existing skills
npx skills update

# Or stay fully current — refresh existing AND pick up any newly-added skills
npx skills add cekura-ai/cekura-skills --all

Remove

npx skills remove cekura-coordinator   # one skill
npx skills remove --all                 # everything

GitHub Copilot: Copilot loads skills from a repository's .github/skills/ directory, so target it explicitly — npx skills add cekura-ai/cekura-skills --all --agent github-copilot --output .github/skills. For the Copilot CLI plugin (skills and MCP tools), see GitHub Copilot below.

What gets installed

Fourteen skills, scoped to specific Cekura workflows:

Skill	When it activates
cekura-coordinator	"What can Cekura do?" — routes you to the right skill
cekura-onboarding	"Get started with Cekura" — full platform walkthrough
cekura-create-agent	"Connect my voice agent to Cekura"
cekura-self-improving-agent	"Improve my agent / auto-tune from eval results" — also "fix a production call bug / reproduce and test a fix before raising a PR"
cekura-metric-design	"Create a metric / measure call quality"
cekura-metric-improvement	"Improve a metric / fix metric accuracy"
cekura-predefined-metrics	"What predefined metrics are available / which built-in metrics should I use"
cekura-eval-design	"Design test scenarios for my voice agent"
cekura-personality-design	"Create a custom personality / test a different accent / make the caller interrupt more"
cekura-infra-test-suite	"Create a committed JSON CI/CD suite / add Tests-as-Code to my voice-agent repo"
cekura-bot-test-writer	"Update the tests for this PR / does this diff need new voice tests / keep my suite in sync"
cekura-flag-call-log-failures	"Analyze the last N calls for issues / what % of calls have "
cekura-generate-scenarios	"Create scenarios from failed calls / regression-test the agent on prod issues"
cekura-results-report	"A weekly status report / a leadership summary / how responsive or secure is the agent? / only failed calls / compare agents / report on this result"

Want full functionality?

npx skills add gives you the behavioral guidance layer — the skills auto-activate when you describe relevant tasks. For slash commands and direct API integration, install the full Claude Code plugin marketplace below.

---

Claude Code (VS Code)

Full plugin support — skills, slash commands, MCP tools, and auto-configured API access.

Install

1. Open the Claude Code chat panel
2. Click Manage Plugins > Marketplaces tab
3. Paste https://github.com/cekura-ai/cekura-skills.git and click Add
4. Switch to the Plugins tab > search for cekura > install the plugin
5. In the Claude Code chat, run /setup-mcp to connect the MCP server. The default is OAuth — a browser opens for one-click sign-in; no API key or restart needed.
6. (Optional) For the API-key auth path or curl-based skill fallbacks, set your key and restart VS Code:

   # Add to ~/.zshrc or ~/.bashrc
   export CEKURA_API_KEY="your-key-here"
   

Get Started

Ask "I'm new to Cekura, help me get started" for a guided walkthrough, or ask "what can Cekura do?" to see everything available.

Upgrade

Run /upgrade-skills in any Claude Code session — it refreshes the marketplace and re-pins the plugin to the latest version, then prompts you to run /reload-plugins (no restart needed in the common case).

To upgrade manually:

claude plugin marketplace update cekura-skills   # refresh catalog from GitHub
claude plugin update cekura@cekura-skills          # move the installed pin to latest

Then run /reload-plugins in your session to apply it. A plain git pull of the marketplace checkout does not move the version pin, so it won't upgrade you on its own — use the commands above.

Auto-update

Installs from this marketplace (cekura@cekura-skills) keep themselves current: the plugin ships a SessionStart hook that runs Claude Code's own update commands at most once per day (best-effort — it never blocks or fails your session, and it no-ops for installs from any other marketplace). To also let Claude Code's native auto-update refresh the marketplace, add this to your ~/.claude/settings.json — or just run /setup-mcp, which offers to enable it for you:

{
  "extraKnownMarketplaces": {
    "cekura-skills": {
      "source": { "source": "github", "repo": "cekura-ai/cekura-skills" },
      "autoUpdate": true
    }
  }
}

---

Claude Code (Terminal CLI)

Same full plugin support as VS Code.

Install

1. Inside a Claude Code session, run /plugins
2. Go to the Marketplaces tab > select Add Marketplace
3. Paste https://github.com/cekura-ai/cekura-skills.git and confirm
4. Go to the Discover tab > search for cekura > install the plugin
5. Run /setup-mcp to connect the MCP server. The default is OAuth — a browser opens for one-click sign-in; no API key or restart needed.
6. (Optional) For the API-key auth path or curl-based skill fallbacks, set your key and restart your terminal + Claude Code session:

   # Add to ~/.zshrc or ~/.bashrc
   export CEKURA_API_KEY="your-key-here"
   

Tip: If you've already cloned the repo locally, you can paste the local path instead of the GitHub URL.

Get Started

Ask "I'm new to Cekura, help me get started" for a guided walkthrough, or ask "what can Cekura do?" to see everything available.

Upgrade

Run /upgrade-skills in any Claude Code session — it refreshes the marketplace and re-pins the plugin to the latest version, then prompts you to run /reload-plugins (no restart needed in the common case).

To upgrade manually:

claude plugin marketplace update cekura-skills   # refresh catalog from GitHub
claude plugin update cekura@cekura-skills          # move the installed pin to latest

Then run /reload-plugins to apply it. A plain git pull of the marketplace checkout does not move the version pin. To pull new versions automatically at launch, see Auto-update (optional) under Claude Code (VS Code), or run /setup-mcp.

---

Codex

Native plugin support — skills and MCP tools. (Codex plugins don't carry slash commands; the skills cover those workflows — they activate by context or when you invoke one with @.)

Install (recommended)

Three steps — add the marketplace, install the plugin, then authenticate the MCP:

# 1. Add the marketplace
codex plugin marketplace add cekura-ai/cekura-skills

# 2. Install the plugin (cekura plugin from the cekura marketplace)
codex plugin add cekura@cekura

# 3. Authenticate the Cekura MCP — opens a browser for OAuth sign-in (no API key)
codex mcp login cekura

Prefer the TUI? Run codex, open /plugins to install cekura, then /mcp to authenticate.

Verify it worked:

codex plugin list   # cekura should be listed as installed
codex mcp list      # cekura should show as connected after login

If you added the marketplace before a recent release, refresh it first so the latest manifest is picked up: codex plugin marketplace upgrade cekura.

API-key alternative: the bundled MCP config is OAuth-only. To use a key instead, add the server to ~/.codex/config.toml with http_headers = { "X-CEKURA-API-KEY" = "your-key" } (or env_http_headers to read it from an env var).

Fallbacks (no MCP)

Skills only — install the skill files directly:

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
  --repo cekura-ai/cekura-skills \
  --path cekura/skills/cekura-onboarding \
         cekura/skills/cekura-create-agent \
         cekura/skills/cekura-self-improving-agent \
         cekura/skills/cekura-metric-design \
         cekura/skills/cekura-metric-improvement \
         cekura/skills/cekura-predefined-metrics \
         cekura/skills/cekura-eval-design \
         cekura/skills/cekura-personality-design \
         cekura/skills/cekura-infra-test-suite \
         cekura/skills/cekura-bot-test-writer

Behavior preset — single-file domain knowledge (metric design, eval design, API reference, anti-patterns):

curl -o AGENTS.md https://raw.githubusercontent.com/cekura-ai/cekura-skills/main/codex/AGENTS.md

Get Started

Ask Codex to help with Cekura metrics or evals — skills load automatically when the conversation matches, or invoke one explicitly with @.

Upgrade

codex plugin marketplace upgrade cekura   # refresh the marketplace snapshot from GitHub
codex plugin add cekura@cekura            # re-pin to the new version (required — not optional)

Both steps are needed: marketplace upgrade only refreshes the snapshot, and codex plugin add re-pins the installed plugin to it. Running just the first leaves you on the old version. Confirm with codex plugin list — the cekura@cekura row should show the new version.

Auto-update (optional). The plugin ships a SessionStart hook that runs the two commands above automatically on each launch (throttled to once/day), so you stay current without typing them. Codex requires you to trust the hook once: run codex, open /hooks, and trust the Cekura SessionStart hook. After that it's automatic; the manual commands above remain available to force an update immediately.

For the fallbacks, re-run the skill installer / re-download AGENTS.md.

---

Cursor

Native plugin support — skills, slash commands, and MCP tools.

Install (recommended)

In Cursor, go to Settings > Plugins > Team Marketplaces > Add Marketplace > Import from Repo, point it at https://github.com/cekura-ai/cekura-skills, then install cekura. Authenticate the Cekura MCP via OAuth when prompted.

Fallback (no MCP)

Use the behavior preset as a rules file:

curl -o .cursor/rules/cekura.md https://raw.githubusercontent.com/cekura-ai/cekura-skills/main/codex/AGENTS.md

Or add it as a global rule in Cursor Settings > Rules.

Get Started

Ask Cursor to help with Cekura metrics or evals — skills load automatically when the conversation matches.

Upgrade

Manual: re-import the marketplace from Settings > Plugins > Team Marketplaces (or re-download the rules-file fallback).

Auto Refresh (optional) — have Cursor pull new versions automatically:

1. Install the Cursor GitHub App on the cekura-ai/cekura-skills repository (Cursor prompts for this; it's required for Auto Refresh).
2. Go to Settings > Plugins > Team Marketplaces and locate the imported Cekura marketplace.
3. Toggle on Auto Refresh and save.

Cursor then re-indexes the marketplace at most once every ~10 minutes, picking up pushed updates to the installed plugin. Note: Auto Refresh updates the existing plugin only — if a brand-new plugin is added to the repo later, re-import the marketplace to pick it up.

Team-wide auto-install (Teams/Enterprise): a workspace admin can mark the Cekura plugin Required for a distribution group in the Team Marketplace settings — that installs and keeps it updated for everyone automatically, no per-developer action. This is an admin dashboard setting, not something the plugin declares.

---

Gemini CLI

Extension support — MCP tools plus the Cekura context file. (Gemini discovers extension skills from a root skills/ directory, so this extension doesn't bundle the nested cekura/skills/; native Gemini skill bundling is deferred.)

Install

gemini extensions install https://github.com/cekura-ai/cekura-skills --auto-update

The --auto-update flag lets the extension self-check GitHub on each launch and apply the new version on the next restart — no manual updates needed. Omit it if you prefer to update manually.

On first use of a Cekura MCP tool, Gemini runs an OAuth sign-in flow. The extension loads GEMINI.md — all Cekura domain knowledge (metric design, eval design, API reference, anti-patterns) — as context.

Upgrade

If you installed with --auto-update, new versions are pulled automatically (restart to apply). To update manually, or to force it now:

gemini extensions update cekura

---

GitHub Copilot

Cekura installs on Copilot two different ways, and they cover different surfaces:

Surface	Install	What you get
Copilot CLI	Marketplace plugin (below)	Skills + MCP tools, OAuth
Coding agent, code review, IDE extensions	Skills committed to .github/skills/	Skills; MCP is configured per repository

Copilot reads its own registry at .github/plugin/marketplace.json, separate from the Claude manifest, so the two installs never interfere. Copilot plugins don't carry Claude-style slash commands — the skills cover those workflows.

Install (Copilot CLI)

1. Install the CLI and sign in (needs a Copilot licence — Pro, Pro+, Business, or Enterprise):

npm install -g @github/copilot

Run copilot once and complete the sign-in prompt.

2. Add the Cekura marketplace:

copilot plugin marketplace add cekura-ai/cekura-skills

3. Confirm Copilot can read it — this should list one plugin, cekura:

copilot plugin marketplace browse cekura-skills

4. Install the plugin:

copilot plugin install cekura@cekura-skills

5. Verify skills and MCP:

copilot plugin list   # cekura listed as installed

copilot mcp list      # cekura listed as an MCP server

6. Authenticate MCP — nothing to configure. The first time Copilot calls a Cekura tool it opens a browser for OAuth sign-in; no API key is stored.

Prefer an interactive session? Run copilot, then /plugin marketplace add cekura-ai/cekura-skills followed by /plugin install cekura@cekura-skills.

Copilot coding agent, code review, and the IDE extensions

Those surfaces load skills from a repository's .github/skills/ directory rather than from CLI plugins. Install the skill files into the repo you want them active in and commit the result:

npx skills add cekura-ai/cekura-skills --all --agent github-copilot --output .github/skills

To give those surfaces Cekura MCP access as well, register https://api.cekura.ai/mcp in your repository's Copilot MCP configuration — see GitHub's docs on configuring MCP servers for your repository. Note that the Copilot coding agent and Copilot code review don't support OAuth-authenticated remote MCP servers, so use the API-key credential there: send your key in an X-CEKURA-API-KEY header. (Copilot CLI does support OAuth — that path needs no key.)

Get Started

Ask Copilot to help with Cekura metrics or evals — skills load automatically when the conversation matches. A good first prompt:

Use the Cekura skills and MCP. List my Cekura agents, pick one, and propose 3 evaluators I should create first. Do not create anything until I approve.

If Copilot answers in generic terms without naming your agents, the skills or the MCP server aren't loaded — re-check steps 3–5.

Upgrade

copilot plugin update cekura

If a release adds a brand-new skill, refresh the marketplace snapshot first: copilot plugin marketplace add cekura-ai/cekura-skills (re-adding refreshes it), then copilot plugin update cekura. For the .github/skills/ path, re-run the npx skills add command above. Copilot has no auto-update hook yet, so run the update command when you want the latest version.

---

Windsurf / Other Agents

Copy codex/AGENTS.md to wherever your agent reads context files from (project root, .windsurf/rules/, etc.):

Install

curl -o AGENTS.md https://raw.githubusercontent.com/cekura-ai/cekura-skills/main/codex/AGENTS.md

The file contains all Cekura domain knowledge in a single portable format that works with any agent.

Upgrade

Re-download:

curl -o AGENTS.md https://raw.githubusercontent.com/cekura-ai/cekura-skills/main/codex/AGENTS.md

---

MCP Server

All plugins connect to the Cekura API through an MCP (Model Context Protocol) server. This gives structured access to 84+ Cekura API operations as typed tools.

For Claude Code users: the plugin auto-configures the MCP server; on first tool use you'll get a one-click browser OAuth sign-in (no API key stored). Run /setup-mcp if tools aren't available — it verifies connectivity and, if you prefer a key-based credential (e.g. for CI), walks you through the X-CEKURA-API-KEY setup.

For Codex, Cursor, Gemini CLI, and GitHub Copilot: the MCP server is wired up natively through each platform's plugin/extension manifest, authenticating via OAuth (no API key stored). The sign-in step differs per platform — in Codex run codex mcp login cekura; Cursor prompts for OAuth when you connect the server; Gemini and Copilot CLI run the OAuth flow on first tool use. For agents using only the AGENTS.md behavior preset (Windsurf, etc.), the MCP server is optional — the preset includes API reference with curl examples as a fallback.

How it works: Claude Code, Codex, and Copilot CLI read the bundled cekura/.mcp.json; Cursor and Gemini declare the endpoint inline in their own manifests. All five point at the Cekura MCP server at https://api.cekura.ai/mcp (CI asserts they stay in sync). By default it authenticates via OAuth — on first use the client opens a browser for a one-click sign-in, with no API key stored. To use an API key instead, run /setup-mcp and choose the API-key path. See the MCP overview.

---

Quick Reference

Slash Commands (Claude Code plugin only)

Command	What it Does
/cekura-onboarding	Guided end-to-end setup — preflight, state-aware resume, walks through agent + metrics + first eval run
/setup-mcp	Configure MCP server (run once after install)
/upgrade-skills	Re-pin the plugin to the latest version, then /reload-plugins
/report-bug	Report a bug — files GitHub issue, optionally attempts a fix
/create-metric	Create or update a metric
/list-metrics	List metrics for an agent or project
/evaluate-calls	Run metrics on specific calls
/improve-metric	Improve metric accuracy: feedback, labs, auto-improve
/autogen-eval	Auto-generate evaluators (or bulk create from CSV/JSON)
/manual-create-update-eval	Create or update a single evaluator with full field walkthrough
/list-evals	List evaluators for an agent or project
/run-evals	Execute test scenarios
/eval-results	Check results from a test run
/cekura-report	Full end-to-end quality report — generates 10 evals, runs them, produces structured analysis
/cekura-results-report	Client-ready report on existing test results, shaped by your question — focus areas, key findings, charts; read-only

Skills (load automatically — both install paths)

Skill	When it activates
cekura-coordinator	"What can Cekura do?" — routes to the right skill
cekura-onboarding	First-time setup, end-to-end platform walkthrough
cekura-create-agent	Setting up an agent — provider, mock tools, KB, dynamic vars
cekura-self-improving-agent	Auto-tuning an agent prompt from eval results — and fixing a production call bug end-to-end (auto-build reproduction harness, must-fail-first gate, fix, stochastic verify, regression sweep, PR / summary)
cekura-metric-design	Designing or creating metrics
cekura-metric-improvement	Improving an existing metric via feedback iteration
cekura-predefined-metrics	Exploring built-in metrics — what each does, costs, constraints
cekura-eval-design	Designing test scenarios for a voice agent
cekura-personality-design	Choosing, forking or creating a personality — the simulated caller's voice, accent, speed, interruption and idle behavior
cekura-infra-test-suite	Creating source-controlled Cekura JSON Tests-as-Code suites for a voice-agent repository
cekura-bot-test-writer	Keeping a committed suite in step with the code — reads a PR diff and makes the smallest edit that closes a real gap
cekura-flag-call-log-failures	Triaging recent production call logs against KPIs — failure rates + outcome distribution
cekura-generate-scenarios	Turning flagged production failures into regression evaluator scenarios
cekura-results-report	Designing a client-ready report on test results that already exist — the ask decides the focus areas, filters and sections (status, leadership, responsiveness, security, P0 flows, failed calls, agent comparison, a topic)

Getting Started Flow

1. /setup-mcp — Configure API access (Claude Code plugin only)
2. /cekura-onboarding — Guided platform setup (preflight + state-aware walkthrough). Or just ask "I'm new to Cekura" to activate the skill directly.
3. Ask "set up my agent" — activates cekura-create-agent
4. /autogen-eval — Auto-generate test scenarios
5. /run-evals — Run your first tests
6. Ask "create a metric for X" — activates cekura-metric-design
7. /cekura-report — Full end-to-end quality report for any agent

---

Platform Compatibility

Platform	Recommended method	Skills	MCP Tools	Slash Commands
Any Agent Skills client	npx skills add	Yes	No	No
Claude Code (VS Code)	Marketplace install	Yes	Yes	Yes
Claude Code (CLI)	/plugins install	Yes	Yes	Yes
Codex	codex plugin marketplace add	Yes	Yes	No
Cursor	Plugin marketplace install	Yes	Yes	Yes
Gemini CLI	gemini extensions install	Context file only	Yes	No
GitHub Copilot (CLI)	copilot plugin marketplace add	Yes	Yes	No
GitHub Copilot (coding agent / code review / IDE)	npx skills add --output .github/skills	Yes	Repo MCP config	No
Windsurf	Rules file	Behavior preset	No	No
Other agents	Copy AGENTS.md	Behavior preset	No	No

---

Reinstalling Cekura skills

If /upgrade-skills reports stale plugin entries, or /plugin install fails with Source path does not exist, your Claude Code install needs a clean reinstall. Run these 4 commands in order in any Claude Code session:

1. /plugin marketplace remove cekura-skills
2. /plugin marketplace add cekura-ai/cekura-skills
3. /plugin marketplace update cekura-skills
4. /plugin install cekura@cekura-skills

Step 3 is the critical one — it forces Claude Code to re-read the marketplace manifest. After step 4, claude plugin list should show one cekura@cekura-skills entry, and all 14 slash commands resolve under cekura:*.

Troubleshooting

/plugin install errors with Source path does not exist: Re-run step 3 (/plugin marketplace update cekura-skills) and retry step 4. If that doesn't clear it, fully restart Claude Code (close and reopen the session), then redo steps 2–4.

claude plugin list shows zero cekura entries after upgrading: you've removed the old plugins but haven't installed the new one. Run steps 2–4 above.

---

Data & privacy

The plugin talks only to your own Cekura account via the Cekura MCP server (https://api.cekura.ai/mcp), with every call mediated by your client's MCP permission system.

• Skill-usage ping: when a Cekura skill or command activates, it calls the cekura_skill_started MCP tool with the skill name, its verification tag, the plugin version, and (for some commands, when available) a conversation/session ID — so Cekura can see which skills are used and validate that the right playbook was loaded. Nothing else from your conversation is sent.
• Local failure log: failures of Cekura MCP tools are logged to ~/.claude/cekura-mcp-failures.log (secret-redacted, capped at 100 lines, never leaves your machine on its own). The /report-bug command may include redacted excerpts in a GitHub issue — only after showing you the full issue body and getting your explicit OK.
• Auto-update: see the Auto-update sections above (Claude Code: daily self-update hook for installs from this marketplace; Codex: daily hook after a one-time /hooks trust).

See Cekura's Privacy Policy and Terms of Service. Questions or issues: [email protected].

---

Links

• Cekura Dashboard: dashboard.cekura.ai
• Sign Up: dashboard.cekura.ai/sign-up
• API Docs: docs.cekura.ai/api-reference
• LLM-friendly Docs: docs.cekura.ai/llms.txt
• Concepts: docs.cekura.ai/documentation/key-concepts