Mivas Run
Walks a consumer through running MIVAS Bench end to end, interview first: what model they are benchmarking, where the model is served (vendor API, or their own endpoint on an inference provider), and where the Docker container runs (local, any Kubernetes, or any container hosting platform). Then a gated ladder: preflight, deploy, three Bluejay smoke calls, harness + benchmark integrity, full industry run. Use when someone says /mivas-run, "run MIVAS myself", "set up mivas-bench", "deploy a harness for MIVAS", "smoke my harness on Bluejay", "benchmark my voice model on healthcare/legal/customer
Install
npx skills add bluejay-ai-dev/mivas-bench --skill mivas-rungit clone --depth 1 https://github.com/bluejay-ai-dev/mivas-bench.git /tmp/mivas-bench
mkdir -p ~/.claude/skills
cp -r /tmp/mivas-bench/.agents/skills/mivas-run ~/.claude/skills/mivas-runPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# MIVAS run (consumer) Opinionated walkthrough. You interview the user, pick for them when they are unsure, and refuse setups that cannot work. Everything uses the public Bluejay REST API with **their** `BLUEJAY_API_KEY`, their inference, their hosting. Nothing internal to Bluejay. ## Step 0: interview. Do not run any command before it is done. Ask with AskUserQuestion, one question at a time, in this order. Bracketed text is the default you pick if they say "you choose" or do not know. Never silently assume anything else. 1. **What model are you benchmarking?** *(a)* a model already wired up — OpenAI Realtime 2.1 / 2.1-mini, GPT-Live 1, Gemini Live, Nova Sonic 2, Grok Voice, Qwen Audio Realtime, Phonic 1, or the LiveKit cascaded baseline → vendor API. *(b)* **your own or an open-weight model** → it needs an inference endpoint you stand up first; only the cascaded harnesses can point at one. **[openai/realtime-2.1]** Map the answer to a `family/runtime` from [HARNESS.md](HARNESS.md); refuse a model with no harness and offer to build one. 2. **Where is that model served?** Vendor API (they hold a provider key) or their own endpoint on an inference provider. Route to [INFERENCE.md](INFERENCE.md) either way; case (b) above is a **separate setup step that must finish and be curl-verified before any hosting work**. 3. **Where will the container run?** Ask before deciding anything else about hosting, and
Details
- Skill name
mivas-run- Repository
- bluejay-ai-dev/mivas-bench
- Path
.agents/skills/mivas-run/SKILL.md- License
- MIT
- Repo updated
- 16 November 2026
FAQ
What does the Mivas Run skill do?
Walks a consumer through running MIVAS Bench end to end, interview first: what model they are benchmarking, where the model is served (vendor API, or their own endpoint on an inference provider), and where the Docker container runs (local, any Kubernetes, or any container hosting platform). Then a gated ladder: preflight, deploy, three Bluejay smoke calls, harness + benchmark integrity, full industry run. Use when someone says /mivas-run, "run MIVAS myself", "set up mivas-bench", "deploy a harness for MIVAS", "smoke my harness on Bluejay", "benchmark my voice model on healthcare/legal/customer
How do I install the Mivas Run skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add bluejay-ai-dev/mivas-bench --skill mivas-run. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Mivas Run popular?
It lives in bluejay-ai-dev/mivas-bench, which has 49 GitHub stars and 0 forks.
Related skills
Mivas Run
49bluejay-ai-dev/mivas-bench
Runs MIVAS Bench end to end from a consumer seat: prerequisites (Bluejay API key, provider key), a harness (existing or self-built), a Kubernetes deployment (local, EKS, Baseten, or any cluster), then a gated ladder of smoke tests, three Bluejay smoke calls, harness + benchmark integrity, and finally a full industry run. Use when someone says /mivas-run, "run MIVAS myself", "set up mivas-bench", "deploy a harness for MIVAS", "smoke my harness on Bluejay", "benchmark my voice model on healthcare/legal/customer-support", or asks how to plug their own voice agent into MIVAS.
Mislead
113Hmbown/Wizards-of-the-Ghosts
Mislead is defensive deception: honeypots, canary tokens, decoy endpoints, fake datasets, and other convincing false presences that reveal who is snooping. The goal is not to trick legitimate users for sport. The goal is to create an instrumented fake that attracts unauthorized curiosity while the real system remains elsewhere. This makes the spell powerful and ethically sharp-edged. It needs explicit boundaries, isolation, and a clear monitoring objective before it should be deployed.
Nondetection
113Hmbown/Wizards-of-the-Ghosts
Use this spell when a system, file, or workflow is leaking more information than it should — metadata, tracking parameters, unnecessary data collection, or identifiable fingerprints.
Mind Blank
113Hmbown/Wizards-of-the-Ghosts
Use this spell when sensitive information must not cross boundaries — between agent sessions, between tasks, between tenants, or between domains with different trust levels.
Scrying
113Hmbown/Wizards-of-the-Ghosts
Use this spell when you need to see what is happening right now on a distant system rather than reading stale logs or cached reports.
Light
113Hmbown/Wizards-of-the-Ghosts
Use this cantrip when a system, process, or codebase has blind spots that need illumination - not detection of hidden things, but creation of visibility where none exists.
Browse the most popular Claude skills, all devops & cloud skills, or write your own with the skills generator.