Kiro can send each LLM request and each tool call to your OpenTelemetry Protocol (OTLP) collector as an event. Each event is one OTLP log record. LLM request events usually arrive within seconds. Tool call events take longer, because the Kiro client reports each call to Kiro first.
| Event | One record for each | Measured by |
|---|---|---|
kiro.chat.message | LLM request that the Kiro service completes | The Kiro service |
kiro.tool.call | Tool call that finished | The Kiro client |
To set up the export, prepare credentials, and then turn on real-time events. To learn how the events relate to each other, see Understanding real-time events.
- Protocol. OTLP logs over HTTP/protobuf, compressed with gzip. gRPC is not supported for events.
- Delivery. Best effort. Kiro retries an event that can't be delivered for up to about 10 minutes, and then drops it. Kiro doesn't send dropped events again. Rarely, an event arrives twice. See Remove duplicates.
- Content. Events carry IDs, counts, and timings.
kiro.chat.messagealso carries the user's email address. Events don't carry prompts, responses, code, or file content. - Record time. The time when the event happened: when the request completed, or when the tool call finished. The observed time is when Kiro sent the event. For an LLM request, that is usually a few seconds after the record time. For a tool call, the gap is larger, and can be up to 45 minutes when the client reports the call late. See kiro.tool.call.
- Availability. Kiro profiles in US East (N. Virginia) and Europe (Frankfurt).
Every record has instrumentation scope kiro.enterprise.events, severity INFO, and the body kiro.enterprise.event. All data is in the attributes. Each record has an eventName, and the attribute kiro.event.name with the same value. Some platforms, such as Dynatrace, don't show eventName, so filter on kiro.event.name. Kiro leaves out attributes that have no value.
These resource attributes appear once for each delivery:
| Resource attribute | Value |
|---|---|
service.name | kiro-enterprise |
kiro.profile.arn | arn:aws:codewhisperer:<region>:<account-id>:profile/<profile-id> |
kiro.profile.id | The Kiro profile ID |
kiro.account.id | The AWS account that holds the Kiro profile |
One record for each LLM request. One user prompt can produce several requests, for example when the agent calls tools.
| Attribute | Type | Presence | Description |
|---|---|---|---|
kiro.message.id | string | Always | The request ID. Unique for each request. Use it to remove duplicates |
kiro.user.id | string | Always | The user's IAM Identity Center user ID. If the profile signs in through an external identity provider, the user ID from that provider |
kiro.user.email | string | When Kiro finds it | The user's email address. With an external identity provider that has no email address for the user, the user name |
kiro.client.type | string | Always | KIRO_IDE, KIRO_CLI, KIRO_WEB, KIRO_CREW, or PLUGIN. Other values can appear as Kiro adds clients |
kiro.client.version | string | When set | The client version from the request's user agent, such as 3.1.0. Not sent when the client sent no version. An updated client reports its new version after it restarts |
kiro.model.name | string | Always | The model, such as auto. Requests on internal models are not sent |
kiro.conversation.id | string | When set | The conversation of the request. A chat from an earlier Kiro CLI engine gets an ID of the form cli_<uuid>_<8 characters> when you open it on the current engine. It gets a new ID each time you open it. Treat each such ID as its own conversation |
kiro.root_conversation_id | string | When set | The top-level conversation of the session. A top-level request has its own kiro.conversation.id. A sub-agent request has its parent session's ID. Not sent by Kiro CLI on the V2 agent engine |
kiro.agent_mode | string | When set | The agent mode of the session. Same values as kiro.tool.agent_mode. A sub-agent request has the mode of its parent session |
kiro.credits.used | double | When set | Credits that the request used |
kiro.credits.overage | double | When set | Overage credits that the request used |
kiro.tool_call_count | integer | When set | Number of tools that the model asked for |
kiro.tool_calls | string | When the model asked for a tool | A JSON object of tool names and counts, such as {"fs_write":1,"read_file":3}. Up to 32 tools, keeping the most called. A call to an MCP tool is listed as tool_call |
kiro.request.subagent | boolean | When set | true for a request that Kiro starts, such as a sub-agent, a summary, or a session title |
kiro.answer.cut_off | boolean | When set | Whether the answer stopped at the model's output limit |
kiro.response.duration_ms | integer | When set | Time in milliseconds from the request to the complete answer |
kiro.subscription.tier | string | When set | PRO, PRO_PLUS, PRO_MAX, POWER, or POOLING |
kiro.usage.limit | integer | When set | The user's credit limit. 999999 for a shared-pool user with no individual limit |
kiro.overage.enabled | boolean | When set | Whether overage is enabled for the user |
kiro.overage.cap | double | When set | The user's overage cap |
One record for each tool call that finished. A request in which the model asks for no tool has no kiro.tool.call record.
The record time is when the tool call finished. If the client reports the call more than 45 minutes late, the record time is when Kiro received it.
| Attribute | Type | Presence | Description |
|---|---|---|---|
kiro.tool.use.id | string | Always | Unique for each tool call. Use it to remove duplicates |
kiro.message.id | string | Always | The request that asked for the tool. All tool calls of one request share it |
kiro.user.id | string | Always | Same value as on kiro.chat.message |
kiro.conversation.id | string | When set | Same value as on kiro.chat.message |
kiro.client.type | string | Always | KIRO_IDE, KIRO_CLI, KIRO_WEB, or UNKNOWN |
kiro.tool.name | string | When set | The tool, such as read_file or invoke_sub_agent. An MCP tool is mcp_<server>_<tool>, with each - in the server name written as _, such as mcp_my_mcp_server_find_city. Kiro web search is remote_web_search. A skill load is disclose_context |
kiro.tool.kind | string | Always | BUILTIN, MCP, SKILL, or UNKNOWN |
kiro.tool.outcome | string | When set | COMPLETED, ERROR, FAULT, ABORT, or UNKNOWN. A call that the user denies or cancels is ABORT. FAILURE appears only when the client sends it |
kiro.tool.duration_ms | integer | When set | How long the call took, in milliseconds |
kiro.tool.agent_mode | string | When set | VIBE, SPEC, QUICK_SPEC, BUG_FIX, PLAN, AUTONOMOUS, CUSTOM, or UNKNOWN. The client shows VIBE as Default |
kiro.error.type | string | When the call didn't complete | The type of error |
kiro.mcp.server | string | MCP calls | The configured MCP server name. remote is Kiro web search, not a server that you configured |
kiro.skill.name | string | Skill loads | The name in the skill's SKILL.md file |
kiro.source | string | Always | CLIENT. The client measures every tool call |
The Kiro client reports tool calls, so some attributes need a supported client version. kiro.chat.message events arrive from every client version.
For Kiro IDE, update to the latest release.
For Kiro CLI, the table lists the minimum version that Kiro supports for each kind of data. Some earlier versions send part of it, but Kiro doesn't support them for this data. Kiro CLI sends tool call data only on the V3 agent engine.
| Data | Kiro CLI |
|---|---|
kiro.tool.call events | 3.0.0 or later |
kiro.tool.call events when Collect metrics per user is off | 3.1.0 or later |
kiro.tool.agent_mode | 3.0.0 or later |
kiro.agent_mode and kiro.skill.name | 3.1.0 or later |
Web search as remote_web_search | 3.1.0 or later |
MCP tool names in kiro.tool.name, such as mcp_my_mcp_server_find_city | 3.2.1 or later |
After an update, quit every open IDE window and kiro-cli process, and start them again. An old process keeps sending data with its old version.
Kiro doesn't notify you when delivery succeeds or fails. Check your collector after users are active.
| Check | Expected |
|---|---|
| Records | One for each LLM request, and one for each finished tool call |
| Resource attribute | service.name = kiro-enterprise |
| Attributes | At least kiro.event.name, kiro.user.id, and kiro.message.id, plus kiro.tool.use.id on tool calls |
| Platform | Where to check |
|---|---|
| Honeycomb | The kiro-enterprise dataset. Query COUNT grouped by eventName |
| Grafana Cloud | Explore, your Loki data source, {service_name="kiro-enterprise"} |
| Dynatrace | The Logs app, filtered on service.name = kiro-enterprise |
| Datadog | Log Explorer, filtered on service:kiro-enterprise |
| Amazon CloudWatch Logs | Logs Insights on your log group: filter eventName = "kiro.chat.message" | stats count(*) by attributes.kiro.model.name |
Platforms store attribute names in different ways:
- Grafana Cloud replaces periods with underscores, so
kiro.model.namebecomeskiro_model_name. Don't add| jsonto a query. The body is plain text. - CloudWatch Logs nests attributes. Use
attributes.kiro.model.nameandresource.attributes.service.name. - Dynatrace doesn't show
eventName. Filter onkiro.event.name.
A retry can deliver an event twice. Remove duplicates on these keys:
kiro.message.idforkiro.chat.messagekiro.tool.use.idforkiro.tool.call. Don't usekiro.message.id, because tool calls share it.
Kiro doesn't guarantee the order of events. Sort by record time, not by arrival.
- Temporary failures are retried. An HTTP 5xx, 408, 425, or 429, a timeout, or a connection error is retried for up to about 10 minutes. Then the events are dropped.
- Configuration errors are not retried. Kiro drops the events at once for any other 4xx, such as 400, 401, 403, or 404. It also drops them for a secret that Kiro can't read, or a malformed header.
- A failing profile pauses. After 5 configuration failures with no successful delivery in between, Kiro stops sending for the profile for about 5 minutes, and drops its events. Then Kiro tries again.
- Private connectivity has no fallback. Until the private connection is ready, deliveries fail and are retried as temporary failures.
- Daily metrics don't change. Dropped events don't affect the daily CSV report or OpenTelemetry metrics.
- Honeycomb. Query the events, for example
SUM(kiro.credits.used)grouped bykiro.model.name. Save the query to a board, or add a trigger. - Grafana Cloud. Use LogQL metric queries, for example
sum by (kiro_model_name) (count_over_time({service_name="kiro-enterprise"} [5m])). Use a recording rule to keep the result as a metric. - Amazon CloudWatch Logs. Use Logs Insights. For metric filters, add a log transformer that copies each field you need into a flat field, because metric filters can't read names with periods. Don't use
kiro.user.idas a metric dimension. Each value is a billed metric, and CloudWatch turns off a metric filter at 1,000 distinct values.
- AWS Secrets Manager and AWS KMS. Kiro reads your events secret, and decrypts it with your key, for each delivery. A delivery carries up to 5 seconds of a profile's LLM request events or of its tool call events, never both. A busy profile usually causes two reads every 5 seconds, one for each event type. A large batch can cause more. Your account pays for these requests.
- Your platform. Your observability platform charges for log ingestion and storage.
OpenTelemetry real-time events