llms.txt
Compare with other sites' llms.txt files →# Raindrop
> Raindrop is the AI observability platform for teams shipping agents: agent monitoring and LLM tracing on real production traffic. It traces every run, catches the silent failures that never throw an error (wrong answers, forgotten context, tool loops, frustrated users), groups them into issues with the traces behind them, and tracks any behavior as a metric with Signals. Before release, Raindrop Simulate runs your agent in a synthetic world and replays each pull request, so you see what breaks before your users do.
Raindrop ingests events and traces from AI products through SDKs (TypeScript, Python, Go, Rust, Swift, Java, browser), an HTTP API, OpenTelemetry, and integrations for most agent frameworks. Plans: Hobby is free (1,000 events/month), Pro is $299/month plus metered events with a 14-day trial, Enterprise is custom. Raindrop is SOC 2 Type II certified.
For AI coding agents: if the user has no Raindrop account, follow https://www.raindrop.ai/auth.md to create one with no browser step (`npx -y create-raindrop`). If they have one, install Raindrop with `curl -fsSL https://raindrop.sh/install | bash -s -- --cloud`, then run `/raindrop-setup` to instrument the project. The hosted MCP server is `https://mcp.raindrop.ai/mcp`. The full docs index is https://www.raindrop.ai/docs/llms.txt and the full docs text is https://www.raindrop.ai/docs/llms-full.txt.
**How Raindrop works**
1. Instrument. Run the installer above, or add an SDK or framework integration. Events go to `https://api.raindrop.ai/v1` with your Raindrop write key (`Authorization: Bearer <write key>`). OpenTelemetry traces go to `https://api.raindrop.ai/v1/traces` over OTLP/HTTP (JSON or protobuf; gRPC is not supported). Your model calls still go straight to your model provider; Raindrop records them, it does not proxy them.
2. Detect. Raindrop reads every event. Signals label each event that matches a behavior you describe in plain language: Raindrop drafts a classifier from the description, you review its matches on your real events, then it runs on every new event and backfills the last 90 days (sampled). Presets cover common failures such as forgetting, user frustration, and task failure. Issue Detection publishes issues when a failure pattern spreads, with severity, status, affected event and user counts, and the events that prove it. Stumbles are single conversations where something went wrong.
3. Investigate. The Triage Agent answers questions about your agent's production behavior in Slack (`@Raindrop`), in the web app, and over MCP, and every answer links to the real conversations behind it. Dashboards chart events and traces with RQL; Users shows each person's conversation history; Datasets keep collections of events and traces.
4. Act and verify. Signal Alerts post to Slack when a signal crosses a threshold. Briefs are scheduled investigations: a Daily Digest, custom briefs on your schedule, and Change Reports that start when a pull request merges and report whether the change behaved as intended in production. Experiments compare signal rates between two cohorts of existing traffic (models, prompts, feature flags, date ranges); they do not run or replay your agent. Cost Tracking shows model spend by model, provider, and project. Self Diagnostics lets your agent report its own failures. Issues can open a Devin coding session with "Fix with Devin".
5. Test before release. Simulate runs your agent in a synthetic world and replays each pull request before merge (see Simulations below).
**Straight answers**
- Do I need to bring my own LLM API key? No. Raindrop runs the models behind Issue Detection, Signals, and the Triage Agent without your OpenAI, Anthropic, or other provider key, and does not charge per token. You only need a Raindrop write key.
- Does Raindrop sit in the request path? No. It is not a proxy or gateway; SDKs and integrations record calls and send them to Raindrop.
- How is it different from traditional monitoring? Error trackers such as Sentry catch exceptions. Agents usually fail without one: a confident wrong answer, a forgotten instruction, a tool loop, a frustrated user. Raindrop reads the conversations and traces to find those.
- Does it support OpenTelemetry? Yes, OTLP/HTTP traces, plus native SDKs for TypeScript, Python, Go, Rust (beta), Swift (beta), Java (beta), and the browser.
- Which frameworks are supported? Vercel AI SDK, OpenAI Agents SDK and Agents API, Claude Agent SDK, Claude Managed Agents, LangChain, LangGraph Deep Agents, Mastra, CrewAI, DSPy, Pydantic AI, Google ADK, Agno, Strands, Temporal, AWS Bedrock, Azure OpenAI, Vertex AI, OpenRouter, TanStack AI, MCP servers, and coding agents (Claude Code, Cursor, OpenCode).
- Can my coding agent use Raindrop? Yes. The hosted MCP server at `https://mcp.raindrop.ai/mcp` (OAuth or API key) queries events, signals, issues, conversations, and Triage, and can create signals.
- What is free? Hobby: 1,000 events per month, 1 custom signal, 15 Triage messages per month, plus tracing, search, notifications, signal presets, MCP, and the Query SDK. No credit card. Ingestion pauses for the rest of the month after 1,000 events.
- What does Pro add? $299/month with a 14-day free trial: Issue Detection, full Stumbles, Experiments, conversation summaries, Memory, data export, 10 custom signals, and 500 Triage messages per month. Events are metered, not capped: $0.003 per event up to 1M, then $0.002.
- What does Enterprise add? Custom limits, SSO/SAML, edge PII redaction, audit logs, warehouse exports, priority support, and SLAs.
- Can I keep PII out of Raindrop? Yes. PII Guard redacts at ingestion so only the redacted text is stored (from $80/month). The TypeScript SDK can also redact common patterns before data leaves your servers.
- Is my data used to train models? No, except custom models used only by your company.
- Is it secure? SOC 2 Type II. Data is encrypted at rest on dedicated AWS infrastructure and in transit with TLS. Projects isolate data for separate products, environments, or teams within one org.
## Products
- [Issue Detection](https://www.raindrop.ai/issue-detection/): Groups recurring agent failures across runs into issues, with the original traces as evidence
- [Signals](https://www.raindrop.ai/signals/): Describe a behavior in plain language and track its rate as a metric across all events
- [Triage Agent](https://www.raindrop.ai/triage-agent/): An agent that investigates your agent's production data in Slack, on the web, and over MCP
- [Change Reports](https://www.raindrop.ai/reports/): Connects to GitHub and reports how each shipped change affected real traffic
- [Workshop](https://www.raindrop.ai/workshop/): Free, open-source local debugger that streams agent traces to your browser and coding agent
## Simulations
Raindrop Simulate tests an agent before it ships. For each run it builds a sealed synthetic world around your agent (synthetic services, data, and state, plus the situations you define), runs your real harness and model inside it, and replays each pull request across simulated events, then posts the results on the PR. The agent reads and writes only synthetic state and makes zero production calls. It uses the instrumentation you already have, with no SDK rewrite. Simulate is in early access with a waitlist.
- [Simulate](https://www.raindrop.ai/simulate/): Product page and waitlist
- [How to test AI agents with simulations](https://www.raindrop.ai/resources/how-to-test-ai-agents-with-simulations/): Talk on replaying real traffic against a change, simulating the world around the agent, and checking that a simulation is faithful
- [Series A announcement](https://www.raindrop.ai/series-a/): First look at Simulate
## Docs
- [Introduction and quickstart](https://www.raindrop.ai/docs/introduction.md): What Raindrop does and the one-command setup
- [TypeScript SDK](https://www.raindrop.ai/docs/sdk/typescript.md): Track events and traces from Node.js and edge runtimes
- [Python SDK](https://www.raindrop.ai/docs/sdk/python.md): Track events and traces from Python apps
- [HTTP API](https://www.raindrop.ai/docs/sdk/http-api.md): Send events from any language without an SDK
- [OpenTelemetry](https://www.raindrop.ai/docs/sdk/opentelemetry.md): Send OTLP/HTTP traces to Raindrop
- [Integrations](https://www.raindrop.ai/docs/integrations/overview.md): Vercel AI SDK, OpenAI Agents, Claude Agent SDK, LangChain, Mastra, Bedrock and more
- [Agent Skills](https://www.raindrop.ai/docs/integrations/skills.md): Skills that let a coding assistant set up and investigate Raindrop
- [MCP server](https://www.raindrop.ai/docs/mcp/overview.md): Query Raindrop data from Claude Code, Cursor, Codex, and other assistants
- [Query API](https://www.raindrop.ai/docs/query-api/sdk.md): TypeScript and Python SDKs for reading Raindrop data
- [Signals docs](https://www.raindrop.ai/docs/platform/signals.md): Create, review, and track signals
- [Issues and Stumbles](https://www.raindrop.ai/docs/platform/issues.md): How Issue Detection works and how it differs from Stumbles
- [Triage Agent docs](https://www.raindrop.ai/docs/platform/triage-agent.md): Ask questions in Slack, on the web, and over MCP
- [Briefs](https://www.raindrop.ai/docs/platform/briefs.md): Daily Digest, custom scheduled briefs, and Change Reports
- [Change Reports docs](https://www.raindrop.ai/docs/platform/change-reports.md): How Raindrop reports on each merged change
- [Experiments](https://www.raindrop.ai/docs/platform/experiments.md): Compare signal rates between cohorts of existing traffic
- [Dashboards](https://www.raindrop.ai/docs/platform/dashboards.md): Saved RQL charts over events and traces
- [Cost Tracking](https://www.raindrop.ai/docs/platform/cost-tracking.md): Model spend by model, provider, and project
- [Self Diagnostics](https://www.raindrop.ai/docs/platform/self-diagnostics.md): Let agents report their own failures
- [Projects](https://www.raindrop.ai/docs/platform/projects.md): Isolate data within an org
- [Workshop docs](https://www.raindrop.ai/docs/workshop/overview.md): Install and use the local debugger
- [Plans and limits](https://www.raindrop.ai/docs/platform/plans.md): What Hobby, Pro, and Enterprise include
- [PII redaction](https://www.raindrop.ai/docs/security/pii-redaction.md): Redact sensitive data before it reaches Raindrop
- [Data security and privacy](https://www.raindrop.ai/docs/security/data-security-and-privacy.md): How Raindrop handles customer data
## Case studies
- [Tolan](https://www.raindrop.ai/case-studies/tolan/)
- [Spiral](https://www.raindrop.ai/case-studies/spiral/)
- [Howie](https://www.raindrop.ai/case-studies/howie/)
- [Unstuck](https://www.raindrop.ai/case-studies/unstuck/)
- [New Computer](https://www.raindrop.ai/case-studies/new-computer/)
- [GC AI](https://www.raindrop.ai/case-studies/gc-ai/)
- [Speak](https://www.raindrop.ai/case-studies/speak/)
## Optional
- [Blog](https://www.raindrop.ai/blog/): Product launches and writing on agents, testing, and monitoring
- [Articles](https://www.raindrop.ai/articles/): Comparisons and guides on agent monitoring and observability tools
- [Resources](https://www.raindrop.ai/resources/): Guides and talks on agent observability and testing
- [SOC 2 Type II](https://www.raindrop.ai/soc-2/): Certification announcement
- [Careers](https://www.raindrop.ai/careers/)
- [App](https://app.raindrop.ai)
- [Talk to sales](https://cal.com/team/raindrop-ai/chat-15-min): sales@raindrop.ai