How-to — MCP in Cursor and Claude Code

Eden's MCP server lets coding agents search production traces, triage issues, rerun the debugger, and open fix PRs — without leaving your IDE.

Prerequisites

1. An Eden org with at least one ingested trace.
2. An org API key (Settings → API Keys in the portal).
3. The Eden platform package installed (see Installation below).

Installation

Three supported paths — all use the same eden-mcp binary:

a) Monorepo checkout (preferred)

pip install -e .
eden-mcp --help   # verify

b) Module invocation (no install needed)

python -m eden.mcp_server

c) Cloud API

Point Cursor MCP at the hosted Eden API with your org API key. No local Eden platform install is required for cloud use.

Environment variables

Variable	Required	Purpose
EDEN_API_KEY	yes	Org API key (sk_eden_…)
EDEN_ORG_ID	yes	Org id (org_…)

Cursor

Create .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "eden": {
      "command": "eden-mcp",
      "args": [],
      "env": {
        "EDEN_API_KEY": "sk_eden_…",
        "EDEN_ORG_ID": "org_…",
        "EDEN_BASE_URL": "https://api.edenobservability.com"
      }
    }
  }
}

If eden-mcp is not on your PATH (e.g. inside a virtualenv), use the module form instead:

{
  "mcpServers": {
    "eden": {
      "command": "python",
      "args": ["-m", "eden.mcp_server"],
      "env": {
        "EDEN_API_KEY": "sk_eden_…",
        "EDEN_ORG_ID": "org_…",
        "EDEN_BASE_URL": "https://api.edenobservability.com"
      }
    }
  }
}

Restart Cursor (or reload MCP servers from settings). Eden appears under MCP Tools in the agent panel.

You can also copy a pre-filled config from Portal → Settings → MCP.

Claude Code

Create .claude/mcp.json. Claude Code expects an explicit "type": "stdio":

{
  "mcpServers": {
    "eden": {
      "type": "stdio",
      "command": "eden-mcp",
      "args": [],
      "env": {
        "EDEN_API_KEY": "sk_eden_…",
        "EDEN_ORG_ID": "org_…",
        "EDEN_BASE_URL": "https://api.edenobservability.com"
      }
    }
  }
}

Available tools

Tool	Scope	What it does
search_traces	read	Search traces by query / agent
get_trace	read	Trace metadata and span tree
get_transcript	read	Step-by-step messages and tool calls
list_checkpoints	read	Debugger checkpoints for a trace
rerun_debugger	debug	Rerun from a checkpoint step
search_issues	read	List agent issues
get_issue	read	Full issue + example stumbles
get_trace_stumbles	read	Stumble detections for a trace
triage_issue	debug	Root-cause triage for an issue
run_autofix	fix	Propose patch + optional PR
get_improvement_queue	read	N17 closed-loop action queue
list_improvement_plans / get_improvement_plan / approve_improvement_plan	read/fix	N23 plans
export_fixture / create_regression_case	read/fix	Failure fixtures + regression cases
preview_change_impact	read	N04 change impact preview
eden_health	read	F84 self-obs health
tail_trace	read	N11 live session tail
label_tool_call	debug	F21 tool-call labels
query_eql	read	Compile / preview EQL filters
query_sql	read	Org-scoped read-only SQL
list_datasets / list_experiments	read	Eval surface mirrors
get_judge_scores	read	Layer A / judge scores for a trace
get_cost_breakdown	read	FinOps explorer rollup
get_agentic_metrics	read	F64 agentic metrics pack

Scopes mcp:read, mcp:debug, and mcp:fix gate tools. Create an MCP-scoped key from Settings → MCP (POST /orgs/{org}/mcp/keys) or use MCP presets on Settings → API Keys.

Streamable HTTP (W28)

In addition to stdio (eden-mcp / python -m eden.mcp_server), Eden exposes JSON-RPC over HTTP:

POST https://api.edenobservability.com/v1/mcp
Authorization: Bearer <org-api-key>
X-Org-Id: org_…

Methods: initialize, tools/list, tools/call, ping. GET on the same path returns a short SSE hello for endpoint discovery. Prefer stdio for Cursor; use HTTP for remote agents.

Example prompts

Try these in Cursor's agent chat after MCP connects:

• "Why did trace abc-123 fail? Show the transcript and any quality signals."
• "List open P0 agent issues from the last 24 hours."
• "Triage issue iss_456 and summarize the root cause."
• "Show F64 agentic metrics for this org."
• "Run a read-only SQL cost breakdown for the last day."

Smoke test

1. Copy config from Settings → MCP (or paste the JSON above).
2. Replace <your-api-key> with a real MCP-scoped key.
3. In Cursor agent chat: "Call search_traces for org {org_id} with limit 5."
4. Optionally: "Call get_agentic_metrics" and "Call query_eql with nl_text errors".
5. You should see JSON tool results.

If the server fails to start, check:

What's next

• Connect any agent — SDK instrumentation tiers
• Agent sessions — session-first tracing (F23)
• Failure fixtures — export redacted traces for local replay