Raindrop integration: operations for AI agents
Raindrop connects to Type.com as an integration. This page lists the 56 operations documented for Claude and Codex in Type.com: 39 read, 13 write, 4 destructive.
About Raindrop
Investigate your AI product with Raindrop, the observability platform for AI agents (raindrop.ai, not the Raindrop.io bookmark manager): ask Raindrop Triage what's wrong, query events, traces, conversations, signals and issues, and manage signals, dashboards, memory, datasets and experiments, through Raindrop's hosted MCP server.
Categories: Monitoring & observability, Developer tools
How to connect Raindrop to Type.com
- Connect
- API key (from Raindrop Settings > API Keys)
- Access
- Organization connection: one shared account for the workspace.
To add the connection, follow the steps in Space connections.
Documented operations
56 operations. Your connection's permissions determine which operations are available.
56 operations listed.
- Read
Get agent conversation
Get the full history of a Raindrop agent conversation including all turns (user messages and agent responses). Use this to review past investigations or to understand context before sending a follow-up via ask_agent_question.
- Read
Get agent progress
Poll the status of a triage agent turn. Each call long-polls for up to 15 seconds. If the turn completes within that window the full response is returned immediately, otherwise returns a progress object with partial_text and completed_tools. IMPORTANT: Print a short status update to the user every time you receive a progress response (e.g. 'Raindrop is investigating... (used: list_issues, search_stumbles, count_events)'), then call again immediately. Never silently poll. When done, relay the response — use ask_agent_question for follow-ups.
- Read
Get application overview
Get a snapshot of your application: event counts with period-over-period trends, recent Issue Detection distribution changes, and top active signals. Pass 'project' to scope the entire snapshot (stats, recent issues, and top signals) to one project. Issues and signals carry url, their shareable app link; use it verbatim when linking.
- Read
Get conversation
Conversation overview: metadata plus slim turns (truncated I/O, event_name, tool counts). For full untruncated turns, use list_events with the convo_id. Paginate via page_info.next_cursor. url is the conversation's shareable app link and each turn's url opens the conversation on that turn; use them verbatim when linking. Example: get_conversation({ conversation_id: "175bc9e1-4e99-44e5-9fa5-3b6d35e29135", project: "triage" })
- Read
Get dashboard
Read a saved dashboard's full definition, panel IDs, queries, visualizations, saved dashboard-wide filters, time settings, and revision before editing. Pass dashboard_id or an exact name and a concrete project. Use list_dashboards for discovery. The built-in Usage board is read through list_dashboards and cannot be edited. For application counts, trends, issues, and signals use get_application_overview (the former get_dashboard tool).
- Read
Get error span facets
Rank failing spans by name and event model. Use this for the Tool Errors tab (failing tools by model).
- Read
Get event
Get ONE event by ID with enrichment: user_traits, per-tool stats, truncated system_prompt_snapshot (or a system_prompt_note if too large) — on top of full input/output. To read all turns of a conversation, use list_events with convo_id instead of looping get_event. url is the event's shareable app link; use it verbatim when linking to the event. Example: get_event({ event_id: "badefa9c-f539-4eb1-aa88-8ec388f90d23", project: "triage" })
- Read
Get event count
Count events matching filters over a recent period, such as the last 24h. Use for a single total without fetching event rows. For an exact historical from/to window, a rate, or grouped distinct counts, use run_rql.
- Read
Get event facets
Get the top values of one event field with counts over a recent period. Use to discover event names, users, models, or triggered signal IDs. Top values are not a complete population. Use run_rql for multiple grouping fields, distinct counts per group, or an exact historical from/to window. Use list_signals for signal names, groups, sentiment, and activity counts.
- Read
Get event timeseries
Get event counts over a recent period, bucketed by the specified interval. Use for a volume trend; use run_rql for an exact historical from/to window, rates, or ranked time buckets. Ensure the period is larger than the interval (e.g. use '7d' or wider for weekly buckets, '30d' for monthly).
- Read
Get experiment
Start here for experiment comparisons: use a known saved name or id, or omit both to list experiments in the selected project, newest first. Duplicate names return matches to select by id. Results include success_rule, evaluation (effective unit and sentiment-derived direction, sample check, 95% two-proportion intervals, p-values), cohort sizes, signal rates, and model_usage. model_usage gives each cohort's per-event input tokens, output tokens, and USD cost as p50/p90, priced like the Costs page over its 2,000 most recent events; cost percentiles skip events with unpriced model calls, and a cohort is null when its usage could not be read. Report the returned verdict; do not calculate a second winner from individual rows or a fixed delta threshold. Explicit selections include hidden and low-performance signals; all mode uses visible active topic, code, and instrumented signals. Selected signals that cannot be evaluated make the verdict unavailable. Metric arrays follow metric_columns; present counts with rates, and distinguish change_pp from relative_change_pct. Both user/event metrics remain available, but evaluation.unit determines the verdict. Follow next_cursor with the same name or id until null. Null rates are undefined. Confidence describes the measured comparison, not causation. Existing experiments support up to 30 days per cohort; new experiments are limited to 14 days.
- Read
Get issue
Get the full AI-generated investigation report for a specific Issue Detection distribution change. Includes the title, description, affected dimensions (the tools/models overrepresented in matched events), and timeline of events. Use this after seeing an issue in get_application_overview or list_issues. url is the issue's shareable app link; use it verbatim when linking to the issue.
- Read
Get issue events
List events matched by an issue. Returns paginated results sorted by most recent. Each event's url is its shareable app link; use it verbatim when linking to the event.
- Read
Get issue timeseries
Get an issue's matched-event counts over time as a timeseries, bucketed by the specified interval.
- Read
Get signal
Get detailed information about a signal including its definition, occurrence count, affected user count, and trend over time. Includes a timeseries (hourly for periods <= 24h, daily otherwise). url is the signal's shareable app link (metric signals have none); use it verbatim when linking.
- Read
Get signal creation status
Check a background signal creation only when the user asks for an update. Use the session_id returned by create_signal and the same project. Call once, report the current status, and return. Scan totals apply to the current classifier revision and reset after refinement, not the whole creation. Report phase and scan totals; progress, when present, describes the current event batch. Its matched count is provisional; scan.matched_total counts confirmed examples. Do not wait, sleep, or poll while it is creating. Only created confirms installation and includes signal_id. This read does not start, advance, or cancel authoring.
- Read
Get signal group
Get a single signal group by its ID. Returns the group details.
- Read
Get trace
OTEL spans for ONE event or trace. Pass event_id or trace_id (one is required; span_id alone is not enough). With only a span_id, look up its trace_id with run_rql first. Narrow with span_type (TOOL_CALL, LLM_GENERATION, SYSTEM_PROMPT) or status ERROR when you know what you are looking for. limit counts matching spans; page with meta.cursor until meta.has_more is false, keeping the same filters. Report only spans in the response. With span_id, include_context adds ancestors and nearby model generations in the same branch (within 24h of the span start). These are temporal neighbors, not proof of causation. Check context.incomplete and its scan limits; null neighbors mean none found in that window. Each span reports input_chars/output_chars (full sizes); a span cut to fit has truncated: true, so re-read that span_id. Example: get_trace({ event_id: "badefa9c-...", project: "triage", span_type: "TOOL_CALL", limit: 50 }). Next page: get_trace({ event_id: "badefa9c-...", project: "triage", span_type: "TOOL_CALL", limit: 50, cursor: "<meta.cursor>" })
- Read
Get user
Get a single user by their ID, including traits, first/last seen timestamps, and event count.
- Read
Get write key
Retrieve the org's write key (RAINDROP_WRITE_KEY) for SDK instrumentation setup. This is the key used for event ingestion — distinct from the org API key.
- Read
List conversations
List conversations for your organization. Returns paginated results sorted by most recent message first. For a conversation's shareable app link, use get_conversation's url. Example: list_conversations({ project: "triage", limit: 20 })
- Read
List dashboards
List Raindrop dashboards the viewer can see (own plus org-visible), including the built-in Usage board (id default-usage). Omit project to list across every project and all-projects boards. Pass project to list that project only; reuse a known project slug or call list_projects when it is unknown. Without dashboard_id or name, the catalog is titles and JSON for the model: id, title, visibility, updated_at, url, project_id, panel_count, and panels as title, kind, and visualization. project_id is null for all-projects boards. No query text and no numbers. The interactive app renders only for a named-board snapshot: pass dashboard_id or name for headline values in the current window. Snapshots apply saved dashboard-wide filters and return them with the board; each query reports skippedDashboardFilters for unsupported kinds. dashboard_id wins if both are set. Name is a case-insensitive exact title. If a board the user names is missing, tell them private boards (not shared with the org) only appear for their owner. Load raindrop_skills topic dashboards to create or edit a dashboard. Read its full definition with get_dashboard before editing. API keys can create and edit org-shared dashboards.
- Read
List datasets
List datasets. Pass dataset_id to inspect one dataset's rows, reference answers, saved event/trace references, and annotations. List results include stable slugs and row counts. Use get_event or get_trace to inspect an individual artifact. Each dataset includes a url to open it in Raindrop.
- Read
List events
List events with filters. Returns paginated FULL event rows (untruncated input/output, model, signals, flags, properties, tools summary). Pass convo_id to read a conversation's turns in full, paging via meta.cursor until meta.has_more is false. Set include_system_prompt for a truncated system_prompt_snapshot; oversized prompts add a system_prompt_note with span lookup guidance. Each event's url is its shareable app link; use it verbatim when linking to the event. Example: list_events({ convo_id: "175bc9e1-4e99-44e5-9fa5-3b6d35e29135", project: "triage" })
- Read
List feature flag keys
List known feature flag names for the organization. Returns each flag key with a type hint: binary (true/false) or variant (multiple values). Use this to discover available flags before filtering events by feature_flags.
- Read
List feature flag values
List distinct values for a specific feature flag key. Use this to see what values a flag takes (e.g. 'true'/'false' for binary, or 'v1'/'v2' for variant flags).
- Read
List issues
List Issue Detection reports for broader changes in the organization's event distribution. This direct catalog does not include one-off Stumbles. For a broad direct investigation such as what issues happened today, also call search_stumbles. Use ask_agent_question when you want Raindrop Triage to investigate both catalogs plus underlying signals and events. Each issue has a title, description, affected dimensions (the tools/models overrepresented in matched events), and related events. Each issue's url is its shareable app link; use it verbatim when linking to the issue.
- Read
List memories
List Raindrop memories visible to the selected project. Organization memories are shared across the organization; users, product, agent, and guidance memories are project-scoped.
- Read
List projects
Discover project slugs in an organization when the project is unknown. Reuse a slug supplied by the user or already resolved in this organization. If several projects fit, ask the user to choose. Pass the returned 'project' value to any read tool's 'project' parameter to scope the call to that project. The 'is_default' field marks the project used when 'project' is omitted. Provide 'org' to list projects for a specific organization; omit it to use your default.
- Read
List signal groups
List signal groups for your organization. Signal groups are collections of related signals organized by theme (e.g. 'errors', 'performance'). Sorted by most recently created.
- Read
List signals
List signals in a single project (the given project, or your default project if omitted), with names, groups, sentiment, and occurrence_count for the selected period and event filters. Sorted by most matches, then newest. Includes configured signals with zero matches. Counts cover topics, instrumented signals, and code signals; regex and metric counts are null. Each signal's url is its shareable app link (metric signals have none); use it verbatim when linking.
- Read
List users
List users for your organization with optional filters. Returns paginated results. Use get_user to fetch full details for a single known user ID.
- Read
Preview dashboard panel query
Validate and preview one exact query for a saved dashboard panel. Load raindrop_skills topic dashboards first. The dashboard supplies timestamp bounds: omit timestamp and end_timestamp from WHERE and HAVING. Pass an explicit preview time_range of at most seven days in the owning project; use at most 24 hours for raw-text filters. For longer saved ranges, preview a bounded subset and disclose it. Preview does not apply dashboard-wide filters. Results validate the query, not the full dashboard's values. Uses the same retention, redaction, timeout, and concurrency controls as run_rql.
- Read
Query cost
Analyze project-scoped LLM spend and token usage. Use summary for totals, breakdown for model/provider attribution, timeseries for hourly/daily trends, and events for paginated recent events with spend. Requires an explicit from/to window of at most 31 days; hourly trends support at most 7 days. Results include pricing coverage, cache-reporting completeness, token breakdowns, and caveats. Unpriced model calls make total_cost_usd incomplete. Breakdowns rank by total cost descending; events are newest-first, not ranked by cost. Use get_event or get_trace to investigate an event. For operation=events, event_filters selects owning Events by user_id, convo_id, event_name, ai_model, properties, user_traits, feature_flags, or signals. Example: event_filters={"convo_id":[{"$eq":"conversation-123"}]}. Event filters do not apply to summary, breakdown, or timeseries. Event pages are not cohort totals. All-projects queries and function attribution are unsupported.
- Read
Raindrop skills
Load the workflow guide before calling its tools, even when the right tool seems obvious. Read it before making the next tool call. Reuse a guide already in context. 'explore': direct investigation, record reads, search, counts, trends, and facets. 'rql': SQL analytics with run_rql, including rates, grouped distinct counts, and exact historical windows. A workflow answered entirely with run_rql needs only this guide. 'rql_reference': additional tables, fields, and functions; load it only when the rql guide lacks what the query needs. 'dashboards': create or edit a saved dashboard directly. 'signals': create or refine a signal when the user requests a change. 'triage': delegate to the hosted Raindrop agent only at the user's explicit request, or continue an existing Triage conversation. Investigate directly with explore otherwise. Omit topic to list available guides.
- Read
Run rql
Run one read-only SQL SELECT over one project's events, traces, users, or conversations. Use for rates, multi-field breakdowns, grouped distinct counts, ranked time buckets, or exact historical event/trace windows that period-based tools cannot express. For a simple recent count, trend, or top values, use get_event_count, get_event_timeseries, or get_event_facets. Load raindrop_skills topic rql before writing SQL. Raw events and traces require a time_range of at most 7 days; it defaults to the last 7 days. RQL supports identity IN subqueries, but no FROM subqueries, joins, unions, or window functions. Results are redacted and limited to 100 rows by default, at most 1000; a limited result is not a population total. An org runs at most 8 queries at once.
- Read
Search docs
Search across the Raindrop docs to find relevant information, code examples, API references, and guides. Use this tool when you need to answer questions about raindrop, find specific documentation, or understand how instrumentation works.
- Read
Search events
Search events by text or regex. Each event's url is its shareable app link; use it verbatim when linking to the event. Text/regex searches over more than 24h without a user_id or convo_id filter can time out; if you get `error: "search_timeout"`, narrow the period or add a filter instead of retrying.
- Read
Search stumbles
Search Stumbles by text or date. Stumbles are one-off unique failure modes from individual interactions: cases where something bad happened once, not broader changes in the event distribution. For broad questions such as what issues happened today, call this alongside list_issues. Returns 50 Stumbles per page sorted by creation time (most recent first). Default time window is `max(24h, 2 x cadence_minutes)` - wide enough that a low-cadence org always covers at least two scan windows. Pass `created_after`/`created_before` to widen or narrow explicitly. The response includes `cadence_minutes` and `last_run_at` so the caller can frame answers as 'as of <last_run_at>' instead of implying real-time. If has_more is true, call again with the next page number to see more. Each Stumble's url is its shareable app link; use it verbatim when linking to the Stumble.
- Write
Add dataset rows
Upload rows that are not Raindrop events or traces (pasted conversations, a customer's own examples, a golden set) together with their annotation. Each row: name, input, output, metadata (string columns), and the standard annotation that create_signal and auto_tune_code_signal use directly: label "match" | "no_match" (what the signal should say about this row) and an optional rationale (why). Example row: {"name": "ticket 1042", "input": "I've asked three times and nobody answers", "output": "Please hold while I check.", "label": "match", "rationale": "the customer is frustrated with the support they got"}. Rows without label but with metadata columns or tags are still usable: the agent maps those annotations itself, but label is more reliable. At least one of name, input, output or metadata is required. Pass id to make re-uploads idempotent. Up to 1000 rows per call.
- Write
Annotate items
Set the standard annotation on existing dataset items: a saved event (target_kind event, event_id), a trace (trace, trace_id) or an uploaded custom row (custom, row_id). label "match" | "no_match" says what the signal should say about the item, and rationale (optional) why; create_signal and auto_tune_code_signal use these directly. Example: {"target": {"target_kind": "custom", "row_id": "ticket-1042"}, "label": "match", "rationale": "the customer is frustrated with the support they got"}. value instead or as well sets the item's single display tag ({label, color}; colors green, red, blue, orange, purple, gray; calling again replaces the tag). Tags are mapped by the agent when no label is set, but label is more reliable.
- Write
Ask agent question
Delegate to Raindrop's hosted Triage agent only when the user explicitly asks for Raindrop Triage or wants to continue an existing Triage conversation. Starts a conversation or sends a follow-up. For a new conversation, pass the user-selected project in project; omission allows reads across all projects. Naming a project only in the message does not scope the conversation. Returns a turn_id; poll get_agent_progress with it as agent_turn_id until done, error, or cancelled. Reuse conversation_id for follow-ups. For dashboard creation and editing, use the direct create_dashboard and edit_dashboard tools with the dashboards skill. Hosted Triage can also save dashboard changes when explicitly delegated; API keys can save org-shared dashboards.
- Write
Create dashboard
Create and immediately save an org-shared dashboard in one concrete project. Supports API keys and OAuth without a separate dashboard write scope. Load raindrop_skills topic dashboards first. Preview every exact panel query in this project before saving; visualization fields must match query output aliases. The server assigns panel IDs and layout. Use edit_dashboard to add panels to an existing dashboard. Returns the saved ID, revision, panel IDs, and scoped URL. Creation is not idempotent: after an uncertain response, check list_dashboards before retrying.
- Write
Create dataset
Create a shared dataset for saved events, traces, and annotations. The dataset includes a url to open it in Raindrop.
- Write
Create experiment
Save a named experiment comparing two cohorts of existing production traffic. Provide the same configurator_config used by Experiments: a date range on each side (1–14 days; prefer 7 or less) plus feature flag, property, model, or other conditions. Define success in module_config.signals.success by choosing the signal set and users or events; direction follows the selected signals' sentiments (negative-only: fewer negatives, positive-only: more positives, mixed: both). Use signal query IDs from query_id in experiment results (signal_id for instrumented signals, id for topics/code). Keep outcomes out of configurator_config cohort filters. Ask when the signal set or unit is unclear; do not ask for direction. This saves a comparison; it does not run or deploy an agent. Use get_experiment with the saved name to read its signals and verdict. Requires the Experiments plan feature and write:experiments for OAuth.
- Write
Create memory
Create a Raindrop memory. Use category organization for org-wide facts; all other categories are saved to the selected project. Guidance memories require applies_to. Concise memories work best: one or two short sentences per fact. When the user's text runs long, suggest a shorter version before saving.
- Write
Create signal
Create a new signal in the background after the user asks to create it. Load raindrop_skills with topic 'signals' first. Use one RQL query to read up to 100 events from the selected project over the last 24 hours. Use those events to understand the requested signal, clarify scope, and create it. Before starting, establish what should count and what should be excluded. Broad labels such as hallucinations or failures need clarification when matching boundaries remain unclear. Ask one or two focused questions and wait for the answers. Preserve explicit scope rules and exclusions; ask before changing them or adding an exception. Ground the question in an inspected event when available: should this event count, and why? Keep intent close to the user's goal, adding only their scope clarifications. Do not expand context or examples into extra rules or a case list. Use the answers to seed positive_event_ids or negative_event_ids. Reuse other inspected examples whose labels are clear. Both lists may be empty; a clear request needs no extra search or labeling round. Raindrop authors, tests, and installs the signal automatically. Call once, acknowledge that creation started, retain the returned session_id, and return to the conversation. Do not wait, sleep, or poll. Use get_signal_creation_status only when the user asks for an update.
- Write
Dismiss stumble
Dismiss a Stumble the user says is wrong (a false positive). It is hidden from the Stumbles page, Home and search_stumbles, and Stumble detection learns to skip similar cases for 90 days. Use the exact Stumble ID from search_stumbles. Only act on an explicit user request, and only on a Stumble the user identified unambiguously: an ID, one already discussed in this conversation, or wording that picks out exactly one Stumble. If nothing matches or several could, list the candidates instead of acting on the closest one. Pass the user's reason if they gave one. For an org API key or a caller with write:memories, the reason and dismissal_reason may update the organization's memory, which shapes future detection: pass only what the user actually said, never your own guess.
- Write
Submit feedback
Submit feedback to the Raindrop team about something that didn't work, is unclear, or could be improved. Use this tool when you encounter issues with Raindrop's documentation, SDK, APIs, or any other aspect of the product. The feedback is sent directly to the Raindrop team.
- Write
Update dataset
Rename a dataset or change its description. Omit a field to keep it; pass description null to clear it.
- Write
Update issue
Update an issue's priority or transition its lifecycle (ignore / resolve / reopen). 'ignore_reason' is required when lifecycle is 'ignore'. On ignore and resolve by an org API key or a caller with write:memories, the ignore_reason and feedback may update the organization's memory, which shapes future issue detection: pass only what the user actually said, never your own guess.
- Write
Update memory
Update a Raindrop memory's text. For guidance memories, pass applies_to to replace which product surfaces use it. Concise memories work best: one or two short sentences per fact.
- Destructive
Delete memory
Delete a Raindrop memory from the selected organization and project scope.
- Destructive
Edit dashboard
Edit and immediately save an existing dashboard. Supports API keys and OAuth without a separate dashboard write scope. Read get_dashboard first and pass its revision as expected_revision. Use the returned panel IDs for actions. Preview every exact new or changed query in the same project before saving; query changes require the complete query and visualization together. Use set_filters with the complete filter list to replace saved dashboard-wide defaults; preserve filters the user did not ask to remove, and pass [] to clear them. Combine filter and panel actions in one edit. Filter-only edits need no query preview. To filter specific panels, change only their queries. Use semantic panel sizes, and tidy_layout only when the user asks to reorganize the layout; never provide grid coordinates. On conflict, reload with get_dashboard and rebuild the edit. Returns the saved revision, affected panel IDs, scoped URL, and filters when explicitly set.
- Destructive
Refine signal
Refine an existing accepted user JavaScript signal, usually after finding false positives during a signal investigation. Reuse the known signal, project, and event examples when the user asks to refine. Direct requests with event labels or a clear comment are also supported; positive examples are optional. Keep comment brief and close to the user's requested change. Do not expand it into a policy, case list, or implementation instructions. Raindrop automatically applies the resulting definition to that same signal in the background. Call once, acknowledge that refinement started, and continue without waiting or polling.
- Destructive
Update dataset members
Add or remove events and whole traces from a dataset in bulk. Additions snapshot the current artifact into the dataset. Duplicate adds and missing removes are safe no-ops.
No operations match.
Not supported
- Raindrop.io bookmarks: this integration is for Raindrop AI observability
- Sending events to Raindrop (use Raindrop's SDKs)
- Organizations where an owner limits API access to owners and admins, unless the key belongs to one of them
Example prompts
- Ask Raindrop what issues happened in our agent today and pull the trace for the worst one.
- How many events matched the tool_failure signal this week, by day?
- Create a signal in the Support project for users reporting they were charged twice.
- What did we spend on LLM calls over the last 7 days, by model?
Get started
- Open Space settings → Connections and search for Raindrop.
- Connect your account and add the connection to the Space.
- Ask for what you need in any thread in that Space.
See How connections work for access and permissions, or browse all integrations.
Reference date: October 28, 2026. Raindrop API docs.