Vocode
Open source Python library for building real time voice agents over phone, Zoom and WebRTC on the builder's choice of speech and language models. Its hosted API is no longer reachable.
Vocode is an open source Python library for building voice agents: real time, streaming conversations with large language models that run over phone calls, Zoom, WebSockets and WebRTC. It is built around three pluggable pieces, a transcriber, an agent and a synthesizer, and it handles streaming audio, endpointing (knowing when a caller has finished speaking), interruptions and conversation speed.
The library ships integrations with speech to text providers such as AssemblyAI, Deepgram, Google Cloud, Microsoft Azure and Whisper, with LLMs from OpenAI and Anthropic, and with text to speech providers including ElevenLabs, Cartesia, Rime, Azure, Google Cloud and AWS Polly. It can also run open source models locally with no internet connection.
Agents take actions mid call through function calling or scripted phrase triggers, and External Actions let an agent push data to and pull data from any API. A self hosted telephony server built on FastAPI and Twilio handles inbound and outbound calls. A LiveKit integration covers WebRTC, a React library connects web clients, and a Sentry integration in beta traces conversations. The library is MIT licensed and installs from PyPI.
Vocode also ran a hosted API for managing phone agents, with numbers, webhooks, vector database retrieval and analytics. That service is no longer reachable. The vocode.dev site redirects to Vocode's GitHub organization and the API's host no longer resolves, though the API documentation is still online. The project is looking for community maintainers.
Vendor details
Canonical URL
https://github.com/vocodedev/vocode-core
Category
Agent infrastructure
Company status
independent
Use cases & customers
Deployment options
In practice
A developer builds a custom phone agent on Vocode, wiring in their chosen speech to text, language model and text to speech providers, and runs the Twilio telephony server themselves for full control of the stack.
A team needs a voice agent with no cloud dependency. Vocode's fully local setup runs open source transcription, language and synthesis models on the team's own machine.
A startup prototypes a voice assistant on Vocode's open source library, using its endpointing and interruption handling so early demos feel natural without building the audio pipeline itself.
Sources & related URLs
Agentic Index coverage score
6.0 / 14 capabilities · 43%
| Integrations & Tool Calling | Full |
|---|---|
|
The open source library supports custom tools. Actions are tools the agent runs mid conversation through function calling, or when it says a configured phrase, and builders create them by subclassing BaseAction. External Actions send a JSON schema payload to any API the builder names and feed the response back to the agent, with request signing. Actions work with ChatGPTAgent. SourceVocode, docs.vocode.dev (Adding Actions to Agents, External Actions, Action Triggers)read 2026-09-27 |
|
| Workflow Orchestration | Partial |
|
Each conversation runs through a fixed pipeline. A StreamingConversation runs a transcriber, an agent and a synthesizer with endpointing, interruption handling and conversation speed settings, and actions run through an ActionsWorker queue. There is no workflow model for sequencing multi step work the builder designs. The LangChain agent option hands response logic to LangChain. SourceVocode, docs.vocode.dev (How it works, Conversation Mechanics, LangChain Agent)read 2026-09-27 |
|
| Knowledge Grounding & RAG | Not documented |
|
The open source library has no retrieval over the customer's content. It covers conversations, providers, telephony, actions and logging, and context reaches the agent through the prompt and action results. A Vector databases feature, in beta, connects an agent to the customer's Pinecone index, but only through the hosted API (vocode_client.agents.update_agent). That host, api.vocode.dev, no longer resolves, so customers cannot use it today. SourceVocode, docs.vocode.dev (llms.txt, Vector databases)read 2026-09-27 |
|
| Human Oversight & Guardrails | Not documented |
|
Actions run when the model calls them or when the agent says a configured phrase. Nothing in the open source library holds an action for a person, hands off to one or limits an action beyond the builder's own code. A warm transfer, in beta, and a HIPAA flag belonged to the hosted API, whose host no longer resolves. The repository's source also includes a transfer call action, which the open source docs do not cover. SourceVocode, docs.vocode.dev (Action Triggers, Adding Actions to Agents, Warm Transfer)read 2026-09-27 |
|
| Security, Identity & Governance | Not documented |
|
The library is MIT licensed code the builder runs, so identity, access and compliance are the builder's to handle. Vocode names no access model and no attestation. External Actions can sign requests with a shared secret, which authenticates an integration call but does not govern who uses the agent. The HIPAA flag belonged to the hosted API, whose host no longer resolves. SourceVocode, docs.vocode.dev (External Actions, HIPAA Compliance) and the vocode-core repository; github.com/vocodedev/vocode-coreread 2026-09-27 |
|
| Observability & Auditability | Partial |
|
The library offers performance spans and event hooks and leaves the record of each call to the builder. A Sentry integration, in beta, sends Vocode's custom spans for each conversation to the builder's Sentry. The Events Manager emits transcript, action and call events for the builder to handle, and a logging guide covers local and production runs. Vocode names no store of what each agent did on each call. SourceVocode, docs.vocode.dev ([Beta] Sentry Integration, Events manager, llms.txt)read 2026-09-27 |
|
| Memory & State Persistence | Partial |
|
State lasts for the length of a call. The agent works from the call's transcript, and each action's result is added to that transcript and shapes later turns. The telephony server keeps each call's configuration in the builder's Redis. Nothing carries into the next call, and Vocode names no memory layer. SourceVocode, docs.vocode.dev (Adding Actions to Agents, Working with phone calls)read 2026-09-27 |
|
| Deployment & Data Residency | Full |
|
The library runs wherever the builder deploys it. It is MIT licensed and installs from PyPI into the builder's own environment, and the telephony server runs under Docker or directly with Python. A fully local setup runs transcription, the LLM and synthesis on the builder's own machine without internet access. The hosted option no longer resolves. SourceVocode, docs.vocode.dev (Fully local conversation, Working with phone calls)read 2026-09-27 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
The library comes with quickstarts, playground scripts and a set of apps in the repository, including a telephony server, a LiveKit app, a Telegram bot, a React client backend, a LangChain agent and a voice RAG demo. These are scaffolds for channels and integrations, and none is an agent built for a job. SourceVocode, docs.vocode.dev (Open Source Quickstarts, Playground) and the vocode-core repository; github.com/vocodedev/vocode-coreread 2026-09-27 |
|
| Triggers & Channel Coverage | Partial |
|
Agents take calls and sessions over phone, web and meeting channels. The self hosted TelephonyServer answers inbound Twilio calls through a webhook and places outbound calls when the builder's script runs OutboundCall. Agents also dial into Zoom, talk over WebSockets from the React client and run over WebRTC through LiveKit. Every session starts with a caller or the builder's code, and Vocode names no schedule, event or queue that wakes an agent. The hosted API also offered call automation, but its host no longer resolves. SourceVocode, docs.vocode.dev (Working with phone calls, Using WebRTC with LiveKit) and the vocode-core READMEread 2026-09-27 |
|
| Model Flexibility & Routing | Full |
|
The builder chooses every provider. Out of the box transcribers (AssemblyAI, Deepgram, Gladia, Google Cloud, Microsoft Azure, Rev.ai, Whisper, whisper.cpp), LLMs (OpenAI, Anthropic) and synthesizers (Rime, Azure, Google Cloud, Play.ht, ElevenLabs, Cartesia, Coqui, AWS Polly and more) are set per conversation in config, open source models run locally, and a custom agent can wrap any text in, text out model. SourceVocode, vocode-core README and docs.vocode.dev (How it works, Create your own AI Agent, Fully local conversation); github.com/vocodedev/vocode-coreread 2026-09-27 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
The SDK is the product. vocode on PyPI is an MIT licensed Python library with extension points for custom agents, actions, transcribers and synthesizers, a FastAPI telephony server, and a React client library. The Node SDK belonged to the hosted API. SourceVocode, vocode-core repository and docs.vocode.dev (Create your own AI Agent, Adding Actions to Agents, React Quickstart); github.com/vocodedev/vocode-coreread 2026-09-27 |
|
| Testing, Debugging & Optimization | Not documented |
|
Playground scripts let a developer try a transcriber, agent or synthesizer by hand. Vocode names no scored test, quality gate or evaluation harness for agent behavior. SourceVocode, docs.vocode.dev (Playground, llms.txt)read 2026-09-27 |
|
| Browser & Computer Use | Not documented |
|
The browser appears only as a client for voice conversations through the React library. Vocode names no browser, desktop or computer control. SourceVocode, docs.vocode.dev (llms.txt, React Quickstart)read 2026-09-27 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Free (open source, MIT)
What is public
The license (MIT) and the package on PyPI are public. Vocode publishes no current price list; the hosted API's tiers went with the service.
Cost watchouts
The library is free, but every call runs on the builder's own transcription, language model, synthesis and telephony accounts, so cost scales with call minutes. With no commits since November 2024, fixes and provider updates fall to the builder.
Variable cost rationale
No platform fee; spend is the builder's own provider usage, which scales with call minutes.
Additional watchouts
The hosted API is no longer reachable, so every deployment is self hosted; the repository's last commit was in November 2024 and its README asks for community maintainers.
Free / trial
Free and open source (MIT)
Key ambiguities
The free library carries no platform fee, so its cost is the provider stack the builder assembles; there is no Vocode price to compare.
Missing data
No hosted offering or support contract is published.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Vocode
The closest documented capability profiles to Vocode among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Deepgram7.5 / 14Adds documented Human Oversight & Guardrails and Security, Identity & Governance
- Beam7.5 / 14Adds documented Security, Identity & Governance and Prebuilt Agents, Templates & Packs
- Circuit & Chisel4.5 / 14Adds documented Human Oversight & Guardrails
- Inworld AI8.5 / 14Adds documented Human Oversight & Guardrails and Security, Identity & Governance, among others
- Cognee7.0 / 14Adds documented Knowledge Grounding & RAG and Security, Identity & Governance
- E2B8.0 / 14Adds documented Security, Identity & Governance and Prebuilt Agents, Templates & Packs, among others
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded