The Cursor Agent CLI provides multiple output formats with the --output-format option when combined with --print. These formats include structured formats for programmatic use (json, stream-json) and a simplified text format for human-readable progress tracking.
The default --output-format is stream-json. This option is only valid when printing (--print) or when print mode is inferred (non-TTY stdout or piped stdin).

​
JSON format

The json output format emits a single JSON object (followed by a newline) when the run completes successfully. Deltas and tool events are not emitted; text is aggregated into the final result. On failure, the process exits with a non-zero code and writes an error message to stderr. No well-formed JSON object is emitted in failure cases.

​
Success response

When successful, the CLI outputs a JSON object with the following structure: 
{
  "type": "result",
  "subtype": "success",
  "is_error": false,
  "duration_ms": 1234,
  "duration_api_ms": 1234,
  "result": "<full assistant text>",
  "session_id": "<uuid>",
  "request_id": "<optional request id>"
}
FieldDescription
typeAlways "result" for terminal results
subtypeAlways "success" for successful completions
is_errorAlways false for successful responses
duration_msTotal execution time in milliseconds
duration_api_msAPI request time in milliseconds (currently equal to duration_ms)
resultComplete assistant response text (concatenation of all text deltas)
session_idUnique session identifier
request_idOptional request identifier (may be omitted)

​
Stream JSON format

The stream-json output format emits newline-delimited JSON (NDJSON). Each line contains a single JSON object representing a real-time event during execution. The stream ends with a terminal result event on success. On failure, the process exits with a non-zero code and the stream may end early without a terminal event; an error message is written to stderr.

​
Event types

​
System initialization

Emitted once at the beginning of each session: 
{
  "type": "system",
  "subtype": "init",
  "apiKeySource": "env|flag|login",
  "cwd": "/absolute/path",
  "session_id": "<uuid>",
  "model": "<model display name>",
  "permissionMode": "default"
}
Future fields like tools and mcp_servers may be added to this event.

​
User message

Contains the user’s input prompt: 
{
  "type": "user",
  "message": {
    "role": "user",
    "content": [{ "type": "text", "text": "<prompt>" }]
  },
  "session_id": "<uuid>"
}

​
Assistant text delta

Emitted multiple times as the assistant generates its response. These events contain incremental text chunks: 
{
  "type": "assistant",
  "message": {
    "role": "assistant",
    "content": [{ "type": "text", "text": "<delta chunk>" }]
  },
  "session_id": "<uuid>"
}
Concatenate all message.content[].text values in order to reconstruct the complete assistant response.

​
Tool call events

Tool calls are tracked with start and completion events: Tool call started: 
{
  "type": "tool_call",
  "subtype": "started",
  "call_id": "<string id>",
  "tool_call": "<json>",
  "session_id": "<uuid>"
}
Tool call completed: 
{
  "type": "tool_call",
  "subtype": "completed",
  "call_id": "<string id>",
  "tool_call": "<json>",
  "session_id": "<uuid>"
}

​
Terminal result

The final event emitted on successful completion: 
{
  "type": "result",
  "subtype": "success",
  "duration_ms": 1234,
  "duration_api_ms": 1234,
  "is_error": false,
  "result": "<full assistant text>",
  "session_id": "<uuid>",
  "request_id": "<optional request id>"
}

​
Example sequence

Here’s a representative NDJSON sequence showing the typical flow of events: 
{"type":"system","subtype":"init","apiKeySource":"login","cwd":"/path","session_id":"...","model":"Claude <ver>","permissionMode":"default"}
{"type":"user","message":{"role":"user","content":[{"type":"text","text":"Explain X"}]},"session_id":"..."}
{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"Sure, "}]},"session_id":"..."}
{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"here are the details..."}]},"session_id":"..."}
{"type":"tool_call","subtype":"started","call_id":"abc","tool_call":{...},"session_id":"..."}
{"type":"tool_call","subtype":"completed","call_id":"abc","tool_call":{...},"session_id":"..."}
{"type":"result","subtype":"success","duration_ms":1234,"duration_api_ms":1234,"is_error":false,"result":"Sure, here are the details...","session_id":"...","request_id":"..."}

​
Text format

The text output format provides a simplified, human-readable stream of agent actions. Instead of detailed JSON events, it outputs concise text descriptions of what the agent is doing in real-time. This format is useful for monitoring agent progress without the overhead of parsing structured data, making it ideal for logging, debugging, or simple progress tracking.

​
Example output

Read file
Edited file
Ran terminal command
Created new file
Each action appears on a new line as the agent performs it, providing immediate feedback on the agent’s progress through the task.

​
Implementation notes

  • Each event is emitted as a single line terminated by \n
  • thinking events are suppressed in print mode and will not appear in either output format
  • Field additions may occur over time in a backward-compatible way (consumers should ignore unknown fields)
  • The stream format provides real-time updates, while the JSON format waits for completion before outputting results