Questions or issues? Contact us at api-support@manus.ai.
Record format
Every audit event is one record with a single schema, delivered two ways:
- Streamed (push): an OpenTelemetry LogRecord sent over OTLP. Attribute keys follow the OpenTelemetry GenAI semantic conventions, with a few Manus extensions.
- Exported (API): one JSON object per line (NDJSON) inside the export archive’s
events.ndjson.
The values are identical; only the field names differ.
Export line (NDJSON)
Each line in events.ndjson is one event:
{
"event_id": "01JABCDE…",
"team_uid": "team_abc",
"user_id": "114504",
"session_uid": "5YX76pz7Dga3yztNVw97Dh",
"event_name": "TOOL_CALL",
"outcome": "SUCCESS",
"occurred_at": "2026-06-09T12:00:00Z",
"metadata": { "eventId": "01JABCDE…", "eventName": "EVENT_NAME_TOOL_CALL", "genAiToolName": "shell_exec" },
"payload": { "gen_ai_tool_call_arguments_json": { "command": "ls" } }
}
- Top-level fields are convenience columns in
snake_case.
metadata is the full record (keys in camelCase; see Metadata fields). Present on every line.payload is present only when the export was created with include_payload=true (keys in snake_case; see Payload).
OTLP envelope (streamed records)
A streamed record is an OTel LogRecord with:
| Part | Value |
|---|
Resource service.name | manus-agent |
Resource tenant.team_uid | Team the event belongs to |
Resource tenant.region | Tenant data-residency region |
| Scope | manus.audit.tier1 (metadata) or manus.audit.tier2 (metadata + payload) |
| Time | Event time (occurred_at) |
| Observed time | Ingestion time (ingested_at) |
| Severity | INFO (9), WARN (13), or ERROR (17) |
| Body | Empty for Tier 1; the payload for Tier 2 |
metadata object, and the matching attribute on a streamed OTLP record:
| Meaning | OTLP attribute | metadata key |
|---|
| Event ID (ULID) | event.id | eventId |
| Schema version | schema.version | schemaVersion |
| Event type | event.name | eventName |
| Outcome | outcome | outcome |
| User ID | user.id | userId |
| Session (task) UID | session.id | sessionUid |
| Request / turn ID | request.id | requestId |
| Source channel | source_channel | sourceChannel |
| Team UID | resource tenant.team_uid | teamUid |
| Tenant namespace | — | tenantNamespace |
| Tenant region | resource tenant.region | tenantRegion |
| Event time | record time | occurredAt |
| Ingestion time | observed time | ingestedAt |
| Severity | severity number / text | severity |
| Tool name | gen_ai.tool.name | genAiToolName |
| Tool call ID | gen_ai.tool.call.id | genAiToolCallId |
| Tool subtype | gen_ai.tool.subtype | genAiToolSubtype |
| Connector name | gen_ai.tool.connector.name | genAiToolConnectorName |
| Connector ID | gen_ai.tool.connector.id | genAiToolConnectorId |
| Connector type | gen_ai.tool.connector.type | genAiToolConnectorType |
| Agent reply kind | agent.reply.kind | agentReplyKind |
| Client address | client.address | clientAddress |
| User agent | user_agent.original | userAgent |
| Geo country | geo.country_iso_code | geoCountry |
| Input bytes | input.bytes | inputBytes |
| Output bytes | output.bytes | outputBytes |
| Message count | message.count | messageCount |
event.name / eventName — one of USER_CHAT, AGENT_REPLY, TOOL_CALL, TOOL_RESULT. Streamed records and the export’s top-level event_name use the short form (TOOL_CALL); the metadata object uses the full proto enum (EVENT_NAME_TOOL_CALL). The same short-vs-full difference applies to outcome (SUCCESS vs OUTCOME_SUCCESS) and agentReplyKind (notify / ask vs AGENT_REPLY_KIND_NOTIFY).request.id — shared by every event in one agent turn; use it to group a user message with the replies and tool calls it triggered.source_channel — how the interaction reached Manus: in_app (the Manus web app) or instant_messaging (an IM integration such as Slack).- Tool attributes — present on
TOOL_CALL / TOOL_RESULT. gen_ai.tool.name is the high-level tool (e.g. shell_exec); gen_ai.tool.subtype is the model-facing name (e.g. shell). The gen_ai.tool.connector.* keys are Manus extensions, present only when the tool is a connector (e.g. Slack) — its name, UUID, and type (mcp, native, …).
agent.reply.kind — on AGENT_REPLY only: notify (a notification) or ask (a question to the user).- Sizes —
inputBytes / outputBytes count the content bytes; messageCount is the number of messages. In metadata, 64-bit integers (inputBytes, outputBytes) are serialized as strings per protobuf JSON convention; messageCount is a number.
- Reserved — LLM request/response fields (
gen_ai.request.model, gen_ai.usage.*, …) are reserved for a future version and are not emitted today.
Payload (Tier 2)
The payload — the payload object in an export (when include_payload=true), or the body of a streamed Tier 2 record — is a JSON object whose fields depend on the event type. Keys are snake_case:
| Field | Event types | Description |
|---|
chat_text | USER_CHAT, AGENT_REPLY | Message content (plain text or a rich-content array, passed through unchanged) |
attachments | USER_CHAT, AGENT_REPLY | Attached files, if any |
agent_reply_kind | AGENT_REPLY | notify or ask |
gen_ai_tool_call_arguments_json | TOOL_CALL | The tool’s input arguments |
gen_ai_tool_call_result_json | TOOL_RESULT | The tool’s output |
gen_ai_tool_call_status | TOOL_RESULT | Execution status of the tool call |
Inside tool arguments and results, values under secret-like keys (containing token, password, secret, api_key, credential, …) are replaced with [REDACTED] before the payload is stored, streamed, or exported.
