Pipecat

Documented custom tool support. Function calling lets the pipeline's LLM call functions the developer registers to reach external services and APIs during a conversation, with handlers written in the agent's own code, and Flows nodes carry their own functions and actions.

SourcePipecat, docs.pipecat.ai (Function Calling, Flows Functions)read 2026-09-27

A workflow model the builder configures. Pipecat Flows structures a conversation as a graph of nodes, each with its own task, tools and transitions, written as declarative flow configs or in code, with state shared across nodes; multiple LLM agents each own their tools and context and hand off control through activation and deactivation; the pipeline itself composes processors frame by frame.

SourcePipecat, docs.pipecat.ai (Pipecat Flows, Multiple LLM Agents, Agent Handoff)read 2026-09-27

Retrieval assembled per turn from services the builder wires in. Pipecat's knowledge retrieval integrations (Moss semantic search injected into the context, Keenable web search over MCP) and function calls to the customer's own stores bring knowledge into each turn. Pipecat itself maintains no index over the customer's content; any index belongs to the partner, and context is assembled per run.

SourcePipecat, docs.pipecat.ai (llms.txt: Knowledge Retrieval services, Context Management)read 2026-09-27

A human handoff and builder defined constraints, no approval step. Daily PSTN supports transferring a call to a person, and Flows limits the LLM to the functions of the current node, a constraint the builder sets. No mechanism holds an agent action for a person's approval. The budget guardrail floe-guard in Pipecat's catalog is a third party service, not Pipecat's own.

SourcePipecat, docs.pipecat.ai (Daily PSTN, Pipecat Flows)read 2026-09-27

A thin access surface and a compliance claim without a named attestation. Pipecat Cloud uses email OTP two factor sign in, organizations whose members hold roles the docs do not enumerate, and REST API keys scoped to organizations and agents. Administrative API calls are logged for a year. HIPAA controls with a BAA and GDPR are stated, and Daily's trust center at trust.daily.co is available on request. No SOC 2 or other certificate is named for Pipecat Cloud.

SourcePipecat, docs.pipecat.ai (Security Guide, HIPAA Compliance, Accounts and Organizations)read 2026-09-27

Run level tracing built into the product. Pipecat's OpenTelemetry tracing visualizes conversation turns and the services each turn called, built in metrics report time to first byte, processing time and LLM and TTS usage per service, TRACE logging records every frame, and Pipecat Cloud keeps agent and session logs readable in the CLI, dashboard and REST API.

SourcePipecat, docs.pipecat.ai (OpenTelemetry Tracing, Metrics, Pipecat Cloud Logging and Observability)read 2026-09-27

Conversation state inside a session. Context aggregators build the conversation history the LLM reads, context summarization compresses older turns in long sessions, and Flows state management shares data across nodes. Long term memory across sessions comes from partner services in Pipecat's catalog (Mem0, Memcode, MemorySync, Synap), whose memory layers belong to those vendors, not Pipecat.

SourcePipecat, docs.pipecat.ai (Context Management, Context Summarization, State Management)read 2026-09-27

Named regions and customer environments. Pipecat Cloud deploys agents to Daily hosted regions the customer picks (us-west, us-east, eu-central, ap-south) for latency and data residency; Pipecat Enterprise runs the same agents in a self hosted region on the customer's own Kubernetes cluster (AWS EKS and Oracle OKE reference architectures) with secrets that never leave it; and the BSD 2-Clause framework self hosts anywhere.

SourcePipecat, docs.pipecat.ai (Pipecat Cloud Regions, Pipecat Enterprise)read 2026-09-27

Example applications the builder adapts. Pipecat publishes complete single agent and multi agent example apps, Flows examples, recipes and a Voice UI Kit; they are code a developer clones and edits, assets the customer assembles rather than packaged agents a buyer adopts.

SourcePipecat, docs.pipecat.ai (Pipecat Examples, Recipes, Flows Examples)read 2026-09-27

Channel coverage is broad. Agents take calls over Daily PSTN and SIP dial-in, Twilio, Telnyx, Plivo and Exotel WebSockets, WhatsApp Business Calling, WebRTC and WebSockets from web, mobile and embedded clients. Dial-out places a call when the customer's application posts to the agent's start endpoint, and Pipecat Cloud webhooks report events outward. No schedule, event or queue that starts an agent on its own is documented.

SourcePipecat, docs.pipecat.ai (Daily PSTN Dial-out, Pipecat Cloud Telephony, WhatsApp Business Calling)read 2026-09-27

The builder chooses every model. Pipecat swaps speech, language and vision services across more than 200 integrated providers, usually one line of code, with the builder's own keys; Pipecat Cloud bills model providers directly to the customer.

SourcePipecat, pipecat.ai (home) and docs.pipecat.ai (service integrations) and daily.co/pricing/pipecat-cloudread 2026-09-27

A documented SDK and API for Pipecat's own platform. The Python framework is the SDK, with a server API reference; client SDKs cover JavaScript, React, React Native, iOS, Android and C++; Pipecat Cloud has a REST API, a Python SDK and a CLI; and an MCP transport serves a Pipecat bot as an MCP server other agents call.

SourcePipecat, docs.pipecat.ai (llms.txt, Pipecat Cloud REST reference, MCP Transport)read 2026-09-27

A built in evaluation harness on the builder's own agent. Pipecat Evals runs scripted scenarios (user turns with expected phrases, function calls, latency budgets or judged checks) and simulated scenarios (an LLM caller with a persona, a goal and a success criterion) against the same pipeline that ships, with eval suites run concurrently from a manifest and a pass or fail signal for each run.

SourcePipecat, docs.pipecat.ai (Pipecat Evals, Eval Suites)read 2026-09-27

Agent driven control of a page, short of a full browser. A UIWorker reads the customer's web app through accessibility snapshots streamed by the Pipecat client and drives it back, scrolling, selecting, filling inputs and clicking, without a person driving each step. It works only on an app that embeds the Pipecat client and implements the command channel. It is not a browser or desktop the agent drives on any site.

SourcePipecat, docs.pipecat.ai (Controlling the UI)read 2026-09-27