MCP server
Ivy's MCP server lets your coding agent read production Lambda traces, errors and payloads. It is a remote server at https://ivytrace.dev/mcp with OAuth sign-in, and it covers all of your workspaces at once.
Add it to your client
Claude Code:
claude mcp add --transport http ivy https://ivytrace.dev/mcpOr commit it for your team in .mcp.json:
{ "mcpServers": { "ivy": { "type": "http", "url": "https://ivytrace.dev/mcp" } } }Claude (web and desktop): Settings → Connectors → Add custom connector, with the URL https://ivytrace.dev/mcp.
Cursor: in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):
{ "mcpServers": { "ivy": { "url": "https://ivytrace.dev/mcp" } } }VS Code: in .vscode/mcp.json:
{ "servers": { "ivy": { "type": "http", "url": "https://ivytrace.dev/mcp" } } }Other clients: any client that supports remote (streamable HTTP) MCP servers with OAuth.
The first time a tool runs, the client opens a browser window to sign in with your Ivy account.
Tools
All tools are read-only. Every tool except list_workspaces takes optional workspace and project (slug or ID); without them it searches every workspace you belong to, and each result says which workspace it came from.
| Tool | Inputs | Returns |
|---|---|---|
list_workspaces |
Your workspaces and their projects | |
find_failing_invocations |
function, time range |
Recent failures, grouped by error |
search_invocations |
function, status, request ID, path, error message | Matching invocations |
list_error_groups |
time range | Error groups with counts and first and last seen, including throttles |
get_trace_summary |
trace_id, detail level |
The trace as a compact text tree |
compare_invocations |
two trace IDs | What differs between a good and a bad run |
render_waterfall |
trace_id |
Interactive waterfall view (MCP Apps), text elsewhere |
render_graph |
trace_id |
Interactive service graph (MCP Apps), text elsewhere |
render_waterfall and render_graph are MCP Apps: in hosts that support them, they show the dashboard's own trace components in the conversation. The view gets its data from the tool result and makes no calls of its own. Hosts without MCP Apps show the text version.
Example prompts
- "Checkout returns 500s since this morning's deploy. Check Ivy."
- "Why is
orders-api-prod-listOrdersslow today? Compare a slow and a fast invocation." - "Show me the waterfall of the last failed
createOrderinvocation." - "Did the fix work? Look for new errors in the last 30 minutes."
Add the Ivy skill to your repo so the agent knows these workflows and which workspace and project belong to the repo.
Self-hosted
Use your own URL, for example claude mcp add --transport http ivy http://localhost:3000/mcp. With OIDC configured, the MCP endpoint is an OAuth protected resource and clients sign in through your provider (which must support dynamic client registration or client ID metadata documents). Without OIDC it is open, like the rest of a self-hosted install.