MCP Server

Manage prompts, triage production traces, and run evaluations from any MCP-capable agent — Claude Code, Claude Desktop, Cursor, and others — without leaving the chat.

Have shell access instead? The CLI is the lighter path — no server connection to maintain, no tool schemas occupying the context window.

Two ways to connect

Hosted (recommended) — no install

Connect directly over HTTP with just an API key. No local package, no build step.

Claude Code (CLI)

claude mcp add --transport http enprompta https://enprompta.com/api/mcp \
  --header "Authorization: Bearer ep_your_api_key"

Any client that supports the MCP Streamable HTTP transport connects the same way — point it at https://enprompta.com/api/mcp with an Authorization: Bearer header; consult that client's docs for its exact remote-server config syntax.

Local (stdio) package

For clients that don't yet speak remote HTTP, or if you prefer a locally-run process:

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "enprompta": {
      "command": "npx",
      "args": ["-y", "@enprompta/mcp-server"],
      "env": { "ENPROMPTA_API_KEY": "ep_your_api_key_here" }
    }
  }
}

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

Cursor — the same mcpServers block in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global). Restart or reload the agent afterward so it picks up the new server.

Authentication & scopes

Get an API key from Dashboard → Settings → API Keys. Each tool call is gated on the key's scopes — a read-only key can list traces and improve prompts but can't score a trace or start a run. On a missing-scope error, add the scope named in the response; it won't retry on its own.

Plan requirements match the REST API: a Pro or Enterprise workspace is required — editor seats get read/write, free viewer seats are read-only. Free accounts can record traces but can't read them through these tools.

Tools

All 13 tools are identical on both transports — the hosted endpoint and the stdio package share the same handlers.

Tool	Scope	What it does
`save_prompt`	`prompts:write`	Save a new version of a prompt with automatic versioning
`get_prompt`	`prompts:read`	Retrieve a prompt by name — latest or a specific version
`get_versions`	`prompts:read`	View a prompt's version history
`evaluate_prompt`	`prompts:read`	Analyze prompt quality with AI-powered feedback
`improve_prompt`	`prompts:read`	Generate improved variants of a prompt
`list_traces`	`traces:read`	List and filter production traces
`get_trace`	`traces:read`	Read one trace in full — input, output, spans
`score_trace`	`traces:write`	Attach a 0–1 score and reasoning to a trace
`aggregate_traces`	`traces:read`	Counts, rates, and averages grouped over traces in one call
`add_traces_to_dataset`	`datasets:write`	Turn failing traces into regression-dataset fixtures
`run_evaluation`	`prompts:write`	Start a test run against a dataset (asynchronous)
`get_eval_results`	`prompts:read`	Read a run's status and pass rate
`verify_setup`	`traces:read`	Confirm tracing actually works — key auths and a trace has arrived

Agent Skill: trace triage & evaluation

Connecting the server gives an agent the tools; it still has to know the playbook — when to score a trace, when to widen a date filter, how to turn a regression into a fixture. The open Agent Skill enprompta packages that judgment: the same production-failure → verified-fix loop shown in the CLI page's example, as reusable instructions any Skills-compatible agent can follow.

Install

npx skills add enprompta/enprompta-skills --skill enprompta

It prefers the CLI when the agent has shell access — cheaper on context — and falls back to these MCP tools otherwise. It activates automatically on tasks involving traces, eval failures, regressions, or "why did the model do X."

Setting up tracing on a new app instead of triaging an existing one? That's a separate skill — enprompta-instrumentation.