Agent Skill: set up Noveum with your coding agent

Install the official Noveum Agent Skill so Claude Code, Cursor, or any coding agent can integrate the SDK, verify traces, run evals, and apply fixes — inside your own environment.

The Noveum Agent Skill packages everything a coding agent needs to set up and operate Noveum end to end — inside your environment, with your credentials. Your code never leaves your machines; the only thing sent to Noveum is the telemetry the noveum-trace SDK emits, exactly as with a manual integration.

The skill walks the agent through seven steps, each gated on a platform-side acceptance check so the agent can't declare success early:

Integrate the SDK — framework-specific recipes for LangChain/LangGraph, CrewAI, LiveKit, Pipecat, and direct provider-SDK apps
Verify trace completeness — a report card that checks what evals actually need (LLM spans, token usage, message content, system prompt, conversation grouping, voice latency telemetry), not just "traces > 0"
Build an eval dataset from real traffic (ETL with AI-generated mappers)
Select scorers and run evaluations — with credit estimates stated before spending
Diagnose with NovaPilot
Backtest fixes as experiments (where AutoFix is enabled)
Apply validated fixes as a reviewable PR, then verify on post-deploy traffic

Install

Claude Code — project-level (recommended for teams; reviewable in your own PR):

git clone https://github.com/Noveum/noveum-skill /tmp/noveum-skill
mkdir -p .claude/skills
cp -r /tmp/noveum-skill/noveum .claude/skills/noveum

Claude Code — personal: copy the noveum/ folder to ~/.claude/skills/noveum.

Claude Enterprise/Team: an admin can upload the noveum/ folder as an organization skill so every seat gets it with zero per-developer setup.

Other agents (Cursor, Codex, Copilot, …): point the agent at noveum/SKILL.md.

Then ask your agent:

Integrate Noveum into this repo and verify traces are flowing. API key is in NOVEUM_API_KEY; use project "my-app".

What the agent needs from you

An API key — create one in your organization's API keys settings (agents cannot create keys, by design). Export it as NOVEUM_API_KEY; the skill never writes it into files.
Your organization slug and a project name (projects auto-create on first trace).

Pair it with the MCP server

The skill works over plain REST, but it's best paired with the Noveum MCP server: connect your agent to https://noveum.ai/api/mcp (OAuth sign-in for URL-only clients, or the same Bearer key) and it gets the full tool surface — ~60 tools, 16 read-first resources, and 20 guided workflow prompts — with schemas that are always current.

Context-safe by design

Noveum payloads can be large (a single span-heavy trace runs to ~145 KB; full reports and dataset exports to megabytes). The skill ships a context-safety discipline — big responses stream to local files via a bundled downloader (with truncation detection) and are inspected selectively, job status is polled on summary endpoints, and nothing large is ever pulled into the agent's context window. Every instruction was validated against live production data.

For security reviewers

The skill is MIT-licensed, dependency-free markdown plus three small stdlib-only Python scripts (connectivity test, trace-completeness report card, and a download-to-file streamer — read-only against the Noveum API except for one test trace). The data-flow statement is at the top of SKILL.md. Vendoring the folder into your repo via your normal PR review is the recommended enterprise install.