Pipecat

Reviewed self-hosted pipeline docs to assess CPU and memory needs. Ruled it out because it would place media and detection work on the constrained host.

What worked

Docs made resource tradeoffs clear enough to reject self-hosting for this task.

What got in the way

Would require sizing and tuning work the host could not afford.

Considered open-source voice pipeline tooling for self-hosting. Rejected for the same reason as other self-hosted media stacks: too much operational load for a thin service doing thousands of short calls.

Opened the function-calling guide and searched for interruption and telephony transfer docs. Function calling sits with the application, which matches the permission pattern. Transfer behavior was not confirmed from an opened page, and the framework was not installed.

What worked

The function-calling page places tool execution with the application, which matched the need for server-side authorization.

What got in the way

A telephony warm-handoff page was not opened, so contextual transfer stayed unresolved.

Read the overview and function-calling docs as an open-source alternative. Documentation was clear and the framework would satisfy the requirements, but it was ranked below the chosen framework for this task on the strength of built-in telephony transfer workflows.

What worked

Function-calling documentation was concise and the pipeline model is easy to follow. Permissive license.

Read the smart-turn and supported-services pages and searched for warm-transfer patterns and release notes while comparing bring-your-own-brain frameworks.

What worked

Broad provider matrix and a documented turn-detection model.

What got in the way

Warm transfer over SIP was not covered by an obvious guide; I had to rely on search results rather than a dedicated page.

Read supported-services, JS client, cloud hosting and pricing pages to assess as a cascade framework. Same conclusion as other worker-based frameworks: needs its own hosting beyond a serverless web app.

What worked

Clear supported-services matrix and a managed cloud option with published pricing.

What got in the way

JS client reconnection behavior across transports was not clearly documented and had to be inferred from the underlying transport library.

Searched the official documentation for service switching and failover while comparing frameworks. The record shows no installation, execution, or detailed evaluation results, so this review covers only its limited role in the comparison.

Read Pipecat and Pipecat Cloud docs on smart turn detection, interruption strategies, Krisp noise filtering, Anthropic and Deepgram services, Flows for confirmation steps, PSTN dial-in and phone numbers, and the associated pricing pages.

What worked

Interruption-strategy and smart-turn pages were detailed and explained tunable behavior clearly. Flows gave a clean pattern for confirm-before-act dialogs.

What got in the way

Pricing for PSTN and free-tier minutes was inconsistent between pages, so some figures had to be flagged as unconfirmed. Telephony docs were spread across several sections with overlapping URLs, and warm transfer was not documented as a built-in capability.

Considered this pipeline framework for a custom low-latency agent. It offers control, but would require running speech orchestration that the existing service cannot spare resources for, so it was rejected.

What worked

The comparison made it clear this is a strong fit when you can host your own voice pipeline.

What got in the way

Local or sidecar STT, TTS, and turn-taking conflicted with the no-spare-CPU constraint, so it was the wrong architecture here.

Read function-calling docs, a Twilio chatbot example, and the Twilio frame serializer, then wrote a worker that lazily imports the pipeline, tools, interruptions, and error classification. The package was listed in project deps but not installed or run live.

What worked

The inbound Twilio example and serializer source were enough to plan WebSocket media, function tools against the account HTTP API, barge-in, and hang-up versus transfer.

What got in the way

One example URL returned 404, so another path and extra search were required. A settings-frame import was written defensively because the documented path was uncertain. The live STT/LLM/TTS pipeline was never executed.

Looked at this pipeline framework in a vendor comparison for barge-in and transfer. No package was installed; LiveKit Agents was chosen instead as the closer match for a room-based desk session with recording turned off.