RESOURCES · DOCS · AGENT STUDIO · CLI REFERENCE

One pipe-friendly binary. Every verb.

The agentstudio CLI is the same surface the desktop and browser editions ride on. JSON output for automation. No desktop runtime required in CI when replay artifacts are already exported.

Talk to AI Labs All docs

INSTRUMENT READOUT · COMMAND COVERAGE

COMMANDS

Nine verbs. One binary.

Each verb does one thing and writes a portable artifact. Pipe them. Combine them. Run them in CI.

COMMANDSUMMARYOUTPUT

agentstudio connect

Open stdio, SSE, HTTP, or WebSocket MCP connections.

↳ session id

agentstudio record

Capture a session into a portable .ast trace.

↳ .ast file

agentstudio replay

Run deterministic replays with mocked tool outputs.

↳ replay report

agentstudio compare

Run one session against multiple provider profiles.

↳ diff report

agentstudio risk-scan

Invoke mcp-risk-linter and attach findings.

↳ SARIF

agentstudio a2a

Validate cards and run contract tests.

↳ JUnit XML

agentstudio providers

Configure provider keys and run smoke checks.

↳ JSON status

agentstudio observe

Start the local OTLP receiver or import OTEL files.

↳ session id

agentstudio export

Write GitHub Action, JUnit, PR comment, trace card, or intake output.

↳ artifact file

FLAGS

The flags that move the artifact.

Flags are scoped per verb. The same flag name keeps the same meaning across verbs to keep automation predictable.

FLAGAPPLIES TODESCRIPTION

--stdio <cmd>connect

Spawn a local stdio MCP server.

--http <url>connect

Connect to an HTTP JSON-RPC MCP endpoint.

--sse <url>connect

Connect to an SSE MCP endpoint.

--ws <url>connect

Connect to a WebSocket MCP endpoint.

--sandboxconnect

Run the server under sandbox-exec or bubblewrap.

--session <name>record

Name the recorded session for trace browser indexing.

--tool <name>record

Pre-fill a tool call before recording.

--json <payload>record

Inline JSON payload for the tool call.

--compare <providers>replay

Comma-separated provider list for cross-model diff.

--mock <path>replay

Use a mocked tool-output fixture for determinism.

--format <kind>risk-scan, a2a

Output format: sarif, junit, json.

--out <path>export, risk-scan

Output path for the artifact.

--github-actionexport

Emit a GitHub Actions workflow YAML.

--junitexport

Emit JUnit XML for CI dashboards.

--pr-commentexport

Emit a PR-comment markdown block.

--trace-cardexport

Emit a portable trace-card markdown.

--intakeexport

Emit a .auraonepkg intake packet preview.

--auth-token <token>observe

Require a bearer token for OTLP receiver writes.

--max-payload-bytes <n>observe

Reject OTLP payloads above the local safety limit.

--rate-limit-per-minute <n>observe

Throttle OTLP receiver writes per client.

--jsonall

Force structured JSON output for automation.

EXAMPLE SESSION

From connect to workflow YAML.

Connect with JSON output, scan the server, and export a GitHub Action that gates future releases.

↳ SHELL

agentstudio connect --http https://example.com/mcp --json
agentstudio risk-scan ./server --format sarif --out mcp-risk.sarif
agentstudio record --session refund --tool refund.create --json '{"id":"INV-9"}'
agentstudio replay refund.ast --compare anthropic,openai --json
agentstudio export refund.ast --github-action --out .github/workflows/agentstudio.yml

CLI REFERENCE

Pipe it. Script it. Keep it.

Run the CLI in CI. Skip the desktop runtime. The exported artifact is the contract.

Talk to AI Labs All docs