Runtypelabs

Runtype umbrella skill for onboarding and routing. Use when the user mentions Runtype, asks what Runtype can build, needs MCP/CLI/dashboard setup, or wants to scope an AI product on Runtype. Route hands-on product builds to runtype-build-product, live account ops to runtype-admin, Persona widgets to runtype-persona, FPO packaging to runtype-templates, SDK/CLI/Marathon work to runtype-sdk-marathon, and agents that run outside Runtype (Flue, Cloudflare Agents SDK, Vercel AI SDK, LangChain, custom loops; OpenTelemetry traces, flue-otel, eval capture from external runs) to runtype-external-agents. Do not use for generic LLM chat or non-AI infrastructure unless Runtype is explicitly part of the request.

Use when designing, building, or reviewing tools that an AI agent will call: MCP server tools, function-calling schemas, agent toolkits, runtime tools, or an API being wrapped for an LLM. Covers classifying a tool (query, command, discovery; sync or async; atomic to orchestrated), the four rules every tool must satisfy, a pre-ship checklist, and an audit procedure for an existing tool set. Routes deeper work to tool-design-interface, tool-design-output, tool-design-errors, tool-design-composition, tool-design-execution, and tool-design-security. Trigger phrases: "design a tool", "tool schema", "MCP tool", "function calling", "agent can't figure out which tool", "review my tools", "why does the agent keep retrying".

Use when operating a live Runtype account through MCP or Code Mode MCP: inspect resources, debug failed flows or agents, read logs and traces, manage products, surfaces, records, schedules, secrets, models, evals, batches, conversations, client tokens, or make safe account mutations. Includes search/execute patterns, read-before-write, validation, pagination, and conservative destructive-change policy.

Use when embedding, deploying, configuring, styling, or debugging Runtype Persona chat widgets, fullscreen AI assistant layouts, chat surfaces, client-token installs, theme tokens, artifacts, tool/reasoning visibility, programmatic widget access, WebMCP page tools, or browser-side local tools. Prefer generate_persona_embed_code and get_persona_theme_reference over hand-written snippets.

Use when creating, validating, exporting, importing, or improving Runtype Full Product Object templates. Covers FPO and FPO template structure, template variables, pending secrets, setup-required auth, validate_product and validate_product_* checks, distributable product packaging, import readiness, and avoiding credential leakage.

Use when designing how a tool fails for an AI agent: error messages that guide recovery, classifying errors as retryable, permanent, needs-user-input, or needs-auth, confirmation requests for ambiguous input, fuzzy-match thresholds, graceful degradation with partial results, and fallback tools. Trigger phrases: "tool error message", "agent keeps retrying", "agent gives up after an error", "ambiguous match", "which user did it mean", "raw 429", "error handling for agent tools", "recovery guidance".

Use when deciding what a tool should return to an AI agent: shaping raw API responses, keeping results token-efficient, cursor pagination, summary-versus-full detail, per-item results for batch calls, GUI links, next-action hints, resource references for large payloads, and canonical data models across a tool set. Trigger phrases: "tool response shape", "response too large", "context window blown by tool output", "pagination for the agent", "what should the tool return", "partial success".

Use when working with the Runtype TypeScript or Python SDK, FlowBuilder, BatchBuilder, EvalBuilder, CLI commands, Marathon long-running agent tasks, playbooks, model fallback, built-in CLI tools, sandboxes, code-first stored/upsert/virtual agents or flows, local tools, hidden parameters, or source-controlled Runtype workflows.

Use when deciding how an agent tool handles identity, credentials, permissions, scope, and state: injecting secrets server-side instead of through the model, permission gates enforced in code, declared OAuth scopes, audit trails, a who-am-i identity anchor, session context, automatic context injection, and context boundaries such as tenant scope or root paths. Trigger phrases: "API key as a tool parameter", "the agent leaked a secret", "prompt says the agent can't delete", "tenant isolation for tools", "who_am_i tool", "audit log for tool calls", "OAuth scopes per tool".

Use when building, designing, validating, or deploying Runtype AI products with agents, flows, tools, surfaces, records, secrets, schedules, evals, orchestration, webhooks, Slack/email/SMS/Telegram/Discord/WhatsApp/iMessage/messaging/chat/API/MCP/A2A/hosted-page surfaces, commerce tools, agent skills, or hosted product generation. Always fetch current MCP build instructions before creating resources.

Use when deciding how an agent tool executes: synchronous versus async job with polling, idempotency keys so retries are safe, transactional boundaries, compensation for cross-system multi-step operations, and timeout boundaries with partial results. Trigger phrases: "long-running tool", "tool times out", "agent retried and created a duplicate", "idempotent tool", "job id and poll", "rollback when a step fails", "async tool for the agent", "saga".

Use when writing or fixing the contract an AI agent sees for a tool: the name, the description, and the parameter schema. Covers descriptions written for model comprehension, enums and ranges instead of free-form strings, smart defaults, natural identifiers resolved internally, exactly-one-of parameter rules, performance hints, flexible input coercion, and marking query versus command tools. Trigger phrases: "tool description", "write the tool schema", "parameter names for the agent", "agent passes the wrong format", "agent picks the wrong tool", "too many required parameters", "JSON schema for an MCP tool".

Use when an agent runs outside Runtype's hosted execution (Flue, the Cloudflare Agents SDK, Vercel AI SDK, LangChain, or a custom loop) and the user wants Runtype's observability, traces, eval capture, Persona chat, surfaces, or the trace-to-eval improvement loop for it. Covers OpenTelemetry/OTLP ingest, @runtypelabs/flue-otel vs @flue/opentelemetry, Flue on Cloudflare Workers, telemetry API keys, registering a runtype-stream or A2A endpoint so Runtype can call the agent, and which features each path unlocks. Not for agents Runtype itself executes.

Use when shaping a set of agent tools rather than one tool: bundling common sequences into task tools, batch variants, preview and dry-run modes, abstraction ladders from raw to intent-level, explicit tool chains, scatter-gather across sources, and how agents discover tools (registry, layered schema exploration, dependency hints, capability search, health checks). Also covers gateways, adapters for legacy APIs, canonical models, and versioning. Trigger phrases: "too many tools", "agent calls the same three tools every time", "bundle these operations", "batch tool", "dry run mode", "tool discovery", "wrap this legacy API", "tool versioning".