RESOURCES · DOCS · AGENT STUDIO · MCP COOKBOOK

Twenty worked MCP debugging recipes.

Each recipe names the server shape, the connection mode, what to inspect, what to record, and what artifact to export. Copyable. Reproducible.

Talk to AI Labs All docs

INSTRUMENT READOUT · COOKBOOK COVERAGE

BASELINE RECIPE

One known server. One recorded tool call.

Start from a known sample server, record a simple tool call, and keep the generated replay artifact under version control.

Recipes are included for Claude Desktop, Cursor, and Cody so users can bring existing MCP configurations into Agent Studio Open without changing their server code.

↳ COMMANDS

agentstudio connect --stdio "npx -y @modelcontextprotocol/server-filesystem ./fixtures"
agentstudio record --session filesystem-smoke --tool list_directory --json '{"path":"."}'
agentstudio replay filesystem-smoke.ast --export github-action

THE 23 SCENARIOS

Every shape we've seen break.

Each scenario lists the transport and the artifact you should keep when the recipe is done.

NO.SCENARIOTRANSPORTARTIFACT

Claude Desktop stdio server migration

stdio↳ trace card

Cursor workspace MCP configuration review

stdio↳ manifest report

Cody-compatible remote HTTP server replay

HTTP↳ replay .ast

Filesystem destructive-path review

stdio↳ risk SARIF

Fetch server timeout and retry trace

stdio↳ trace card

SQLite read-only query guard

stdio↳ policy note

Memory server prompt-injection probe

stdio↳ risk SARIF

Brave Search rate-limit replay

HTTP↳ replay .ast

OAuth bearer header configuration

HTTP↳ connection profile

mTLS remote server certificate note

HTTP↳ connection profile

SSE reconnect and dropped event recovery

SSE↳ trace card

HTTP JSON-RPC batch failure

HTTP↳ JUnit report

WebSocket session close handling

WebSocket↳ trace card

Tool schema narrowing after linter warning

any↳ manifest diff

Resource listing with empty-state verification

any↳ trace card

Prompt template safety review

any↳ risk SARIF

Server logs attached to trace card

stdio↳ trace card

Mocked tool outputs for replay

any↳ replay .ast

Model upgrade comparison

any↳ compare report

GitHub Action export

any↳ workflow YAML

JUnit export for CI dashboards

any↳ JUnit XML

PR comment export

any↳ PR markdown

AuraOne intake packet preview

any↳ .auraonepkg

TRANSPORT RECIPES

One pattern per transport.

STDIO

Local filesystem server

Spin up the reference filesystem server, scan tools for destructive paths, and capture a list_directory replay.

agentstudio connect --stdio "npx -y @modelcontextprotocol/server-filesystem ./fixtures"
agentstudio risk-scan ./fixtures --fail-on high
agentstudio record --session fs-list --tool list_directory --json '{"path":"."}'

SSE / HTTP

Remote MCP with OAuth

Configure a bearer header, replay a tool call against the remote, and export a PR comment.

agentstudio connect --http https://example.com/mcp --header "Authorization: Bearer $TOKEN"
agentstudio record --session refund-tool --tool refund.create --json '{"id":"INV-9"}'
agentstudio export refund-tool.ast --pr-comment

WEBSOCKET

Streaming session close

Record a long-running tool call across a WebSocket transport and verify a clean close in the trace.

agentstudio connect --ws wss://example.com/mcp
agentstudio record --session stream-close --tool stream.subscribe
agentstudio replay stream-close.ast --assert-close-code 1000

MCP COOKBOOK

One recipe at a time. One escape caught.

Pick the recipe closest to the failure you saw last week. Run it. Keep the artifact.

Talk to AI Labs All docs