Noveum.ai
MCP

Model Context Protocol

Drive Noveum from any AI agent.

Claude, Cursor, VS Code, ChatGPT — one MCP endpoint connects them all to the full Noveum platform. Your agent searches traces, runs evals, launches simulations, and ships fixes, in plain language.

Get startedRead the docs
noveum-mcp
HTTP · OAuth 2.1

MCP endpoint

Tools~60

tracesdatasetsevals

Resources16

noveum://projectsnoveum://scorers

Prompts20

/debug_traces/run_evaluation

One endpoint · every Noveum capability

Connect

Two minutes to connected.

Add one endpoint to the client you already use. Paste an API key for headless setups — or add just the URL and sign in with OAuth.

  1. Add the endpoint

    Point your MCP client at noveum.ai/api/mcp — Cursor, Claude Code, VS Code, ChatGPT, or anything that speaks Streamable HTTP.

  2. Sign in — or paste a key

    URL-only clients open a Noveum sign-in and consent screen automatically. Headless setups send an API key in the Authorization header instead.

  3. Ask your agent

    Tools, resources, and prompts load on connect. Ask for failing traces, an eval run, or a NovaPilot analysis — in plain language.

No key required: clients that support the MCP authorization spec discover OAuth automatically. Add the URL, approve the consent screen once, and you're connected.

Read the full server reference
.cursor/mcp.json
{
  "mcpServers": {
    "noveum": {
      "url": "https://noveum.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer NOVEUM_API_KEY"
      }
    }
  }
}

Tools · Resources · Prompts

Everything the app can do, your agent can do.

Tools are generated from the live OpenAPI spec, resources feed your agent real ids and slugs before it plans, and prompts package entire workflows into a single slash command.

Tools

~60 HTTP actions generated from the live OpenAPI spec — list things, start jobs, poll progress until terminal.

Traces

Search spans, errors, and latency across every deploy.

Datasets

Create, version, and publish eval datasets from real traffic.

Evals

Kick off eval jobs, poll runs, and read scored results.

ETL pipelines

Build extraction pipelines and generate mappers with NovaEval.

Scorers

Browse 100+ built-in scorers or get recommendations per dataset.

Projects

List projects, environments, usage, and organization status.

NovaPilot

Start analyses, read reports, and manage recurring schedules.

NovaSynth

Generate personas, launch simulations, and pull analytics.

Resources

Read-first context served as JSON — so every id, slug, and filter enum comes from real data, not guesses.

  • noveum://projects
  • noveum://scorers
  • noveum://filter-values
  • noveum://org-status
  • noveum://datasets
  • noveum://eval-jobs
  • noveum://etl-jobs
  • noveum://novapilot-schedules
  • noveum://novasynth-runs
  • noveum://agent-versions
  • +6 more

Prompts

Slash-command workflows that spell out which tools to chain, what to poll, and when to stop.

  • /debug_traces
  • /run_evaluation
  • /analyze_with_novapilot
  • /optimize_costs
  • /trace_error_triage
  • /compare_eval_runs
  • /recommend_scorers_for_dataset
  • /full_agent_improvement_cycle
  • /test_voice_agent
  • /download_novapilot_report_markdown
  • +10 more

Recipes

Real workflows, in plain language.

No dashboards, no query language. Point your agent at Noveum and describe the outcome you want — these are things teams run every day.

Find and triage failures

/debug_traces

Let your agent surface what's breaking in production and why, straight from your live traces.

Find the error traces from the last 24 hours, group them by root cause, and show me the three worst offenders with example spans.

Turn traces into an evaluation

/run_evaluation

Build a dataset from real traffic and score model candidates side by side with Noveum's calibrated judges.

Create a dataset from yesterday's failing traces, then run an evaluation comparing gpt-4.1 and gemini-3-flash and tell me which wins on accuracy and cost.

Cut model spend

/optimize_costs

Have your agent read a week of usage and recommend a cheaper setup that holds quality.

Analyze my token usage and latency this week, then recommend a cheaper model setup that keeps quality above 90%.

Close the loop end to end

/full_agent_improvement_cycle

Run the whole trace to dataset to eval to NovaPilot fix to redeploy loop from a single prompt.

Run the full agent improvement cycle on my checkout assistant: trace it, build a dataset, evaluate, have NovaPilot draft a fix, and summarize the pull request.

Authentication

Sign in like a person. Or key in like a machine.

Two modes, one endpoint. Both respect your organization's roles — a connected agent can never do more than you can.

Default

OAuth 2.1 sign-in

For interactive clients. Add the URL, approve a consent screen once, and you're connected — no key to paste, no OAuth app to register.

  • Per-user consent with exactly the scopes you choose
  • Each token is bound to a single organization, picked at consent
  • Revoke any app instantly under Settings → Connected apps
Headless

Bearer API key

For CI pipelines, cron jobs, and server-to-server setups — any client without a browser to complete a sign-in.

  • The same key as the REST API, sent as an Authorization header
  • Scoped to your organization and rotatable at any time
  • Pin a specific key for scripted, reproducible setups

noveum.read

Read projects, traces, datasets, evals, and reports

noveum.write

Create and update datasets, jobs, and settings

noveum.execute

Start jobs — ETL, evals, NovaPilot, NovaSynth

Scopes only ever narrow what a connected app can do. Your role and RBAC permissions in the organization still apply on top.

Works with

If it speaks MCP, it speaks Noveum.

FAQ

MCP questions, answered.

What is the Noveum MCP server?
It's a remote Model Context Protocol server at noveum.ai/api/mcp. Connecting it gives any MCP-compatible AI client — Claude, Cursor, VS Code, ChatGPT, and more — secure, tool-level access to your Noveum workspace: traces, datasets, evaluations, scorers, and agent-improvement workflows, all driven in natural language.
Which AI clients can connect?
Any client that speaks the MCP Streamable HTTP transport — Claude and Claude Code, Cursor, VS Code, ChatGPT (developer mode), Windsurf, Cline, Replit, Zed, Goose, and others. URL-only clients sign in with OAuth; headless or scripted clients like cURL, Python, and CI use a Bearer API key.
Do I need an API key, or does it use OAuth?
Both work on the same endpoint. Interactive clients discover OAuth 2.1 automatically — you add the URL, approve a consent screen once, and you're connected with no key to paste. Headless setups send an Authorization: Bearer header with a Noveum API key instead.
What can an agent actually do through it?
Around 60 tools generated from Noveum's live API — search and read traces, create datasets, run evaluations, manage scorers, and kick off ETL, NovaPilot, and NovaSynth jobs — plus 16 read-first resources and 20 guided workflow prompts. Every action respects your organization's roles and permissions.
Is it secure? How do permissions and revocation work?
OAuth tokens are bound to a single organization you pick at consent time and to coarse read / write / execute scopes that only ever narrow what a connected app can do — your RBAC permissions still apply on top. Revoke any connected app anytime from Settings, Connected apps, and rotate API keys from your org settings.
Does connecting cost anything extra?
No — the MCP server is included with Noveum and there's no separate charge to connect. Usage of the underlying platform (traces, evaluations, and jobs) follows your normal Noveum plan.

Next step

Put your production AI under control.

Start free and ship your first trace in 15 minutes — or book 30 minutes and we’ll integrate live on the call: your stack, your data, your first eval report before it ends.

Start freeBook a demo

SOC 2 Type II (in progress) · GDPR · On-prem & BYO ClickHouse available