> ## Documentation Index > Fetch the complete documentation index at: https://unmute.ai/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > Unmute compiles to exactly four targets. Pipecat and LiveKit are code targets: compile writes a Python project you run. SLNG is a hosted target: compile writes a deployment body and SLNG runs the agent, so it has no `unmute dev`. Twilio compiles to a small Python app that Twilio ConversationRelay calls; you host it, and it has no `unmute dev` either. Those four are the only values `provider` accepts in `targets.yaml`. Deepgram and ElevenLabs appear in these docs as model vendors, which is not the same thing as a target, and `slng` is both. > The Go structs in `internal/spec` and `internal/ir` are the schema truth. Check a field against them, or run `unmute validate`, rather than against what you remember. # Why Unmute Core > What the library adds to a LiveKit or Pipecat agent, and when you need it. LiveKit and Pipecat move audio, run the models and call your tools. They do not decide how a good phone call behaves. Every team learns that from its first real calls. unmute-core holds what we learned, so your agent starts from it. You keep your framework, your plugins and your pipeline. The library adds one piece or all of them. ## One call, two ways A caller phones a salon to book a haircut. Here is what goes wrong on a plain agent, and what the library does instead. | The call | A plain agent | With unmute-core | | - | - | - | | The caller dials in and talks while the agent sets up. | That audio mixes into the first turn. | Caller audio waits until the greeting starts. | | "My number is six one two, three four five..." | The model saves `612345` and books with it. | The save is refused because the number is not valid. The model asks again. | | The caller pauses, then goes on. | The reply plays over them. | The reply is dropped. The next one answers everything they said. | | "Yes" while the agent reads the times. | The agent stops mid-sentence. | The agent finishes, then answers the "yes". | | The model says "Noted, thank you." | The step waits for a finish that never comes. | The library asks the model directly: finished, or not yet? | | The booking tool takes 3 seconds. | Silence. | "One moment, let me check." | | The model calls the booking tool. | It books what the model guessed. | Code asks "Shall I book Friday, October 16?" and books only what the caller heard. | | The booking API answers with half a record. | The model tells the caller it is booked. | The answer is checked against the tool's model. The model reads which field is missing and does not claim success. | | The caller is handed to billing. | Billing hands the call straight back. | Billing speaks first. Handoff tools wait for the caller. | | The model provider fails. | Silence, then the line drops. | The next model answers. If none can, a short apology plays. | | The call ends. | Each framework gives a different reason, and its own trace shape. | One reason and one trace on both frameworks. | Each row is a few lines of code to fix, once you know it happens. Most teams find out on a live call, one row at a time. ## Three reasons to use it 1. **One agent on both frameworks.** State, tasks, tools, prompts and traces behave the same on LiveKit and Pipecat. 90 shared test scenarios check it on every change. You can switch frameworks without rewriting the agent. 2. **The model cannot break your data.** Each value has a type. A save that does not fit is refused with a reason the model can act on. Values the caller has not confirmed stay out of the prompt. 3. **What real calls taught us, built in.** Pauses, interruptions, idle callers, slow tools, failing providers, transfers and handoffs. Each feature page starts with what it fixes. ## Where each part helps | You want to | Read | | - | - | | Collect values the caller gives, and trust them | [State](/core/build/state) | | Walk the caller through steps that cannot get stuck | [Tasks](/core/build/tasks) | | Call your code and APIs safely during a call | [Tools](/core/build/tools) | | Pass the call between agents | [Handoffs](/core/build/handoffs) | | Open, fail and end a call well | [Call lifecycle](/core/build/call-lifecycle) | | Know the caller before the first word | [Prefetch](/core/build/prefetch) | | Get pauses, interruptions and quiet callers right | [Turn taking](/core/build/turn-taking) | | Change prompts, requests and tool calls in one place | [Hooks](/core/build/hooks) | | Answer from your documents | [Knowledge](/core/build/knowledge) | | Keep answering when a model fails | [Model call](/core/models/model-call) | | Transfer to a person, dial out, handle voicemail | [Telephony](/core/telephony) | | See every call in Langfuse, Coval, Logfire or Datadog | [Telemetry](/core/observe/telemetry) | | Spend fewer tokens, and say the same words on every call | [Type the path](/core/examples/typed-path) | | See whole agents built this way | [Example agents](/core/examples/overview) | ## When you don't need it * One agent on one framework, with a few tools and no values to collect. * You already built and tested your own answers to the rows above. Even then, you can take one piece, like [tools](/core/build/tools) or [telemetry](/core/observe/telemetry), and add more when a call shows you need it. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.