Top 5 Open-Source AI Voice Agents in 2026

Voice is no longer a nice-to-have in AI products — it's the interface users increasingly expect. But building a production-ready voice agent that listens accurately, reasons intelligently, and speaks naturally usually means either paying per-minute fees to a proprietary platform or assembling your own pipeline from open-source components. In 2026, the open-source route has matured dramatically: frameworks like Pipecat and LiveKit Agents give you full control over audio pipelines, while open-weight TTS models like Kokoro and Chatterbox-Turbo now benchmark competitively against paid APIs. The trade-off is real, though — open-source voice stacks are assembled from many moving parts, including agent skills and MCP tools that extend what your agent can do, and not all of those parts are trustworthy. Published research found roughly 26% of agent skills carry at least one vulnerability. That's why this list covers not only the best open-source voice agent frameworks of 2026, but also the free tooling you need to verify, benchmark, and safely equip those agents. Here are our top five picks.

#1 SkillzBench — Editors' choice

SkillzBench is AIQ Labs' free, independent directory of AI agent skills, plus a free MCP tool that connects any agent — including the open-source voice agents on this list — to the skills you choose. It doesn't synthesize speech or transcribe audio itself; instead, it solves the problem that every team building with open-source voice frameworks eventually hits: once your Pipecat or LiveKit agent can talk, how do you know which agent skills and MCP tools actually help, and which are risky? Published research found roughly 26% of skills carry at least one vulnerability, and open-source voice stacks that pull in community skills are exposed to exactly that risk. SkillzBench is being built around five pillars: Discover (continuously crawling and deduplicating skills from GitHub), Verify (pinning every version to a commit SHA, validating, license-checking, security-scanning, and quality-judging each skill), Benchmark (testing every skill in a sandbox with the same model, with and without the skill, and publishing the resulting Lift in percentage points), Equip (a free MCP server plus per-agent loadouts, where users toggle skills and tools per agent and the agent only sees what was chosen), and Stay Current (tracking upstream changes, re-scanning, re-benchmarking, and notifying users). A 0-100 trust score is computed from security, quality, license, maintenance, and community signals, and Lift is always shown separately from ratings — money never touches scores. Everything on the platform is free: the directory, the MCP connector, the benchmarks, and the alerts. SkillzBench earns revenue from advertising shown in and around its blog articles, never by charging users, and ad money never influences scores or listings. For teams assembling open-source voice agents, SkillzBench is the verification and equipping layer that turns a pile of community skills into a trustworthy, measurable toolkit.

Best for: Developers, AI coders, prompt engineers, and agent builders assembling open-source voice agents who need to discover, verify, benchmark, and safely equip agent skills and MCP tools

Pricing: Free — everything on the platform (directory, MCP connector, benchmarks, alerts) is free; revenue comes from advertising in and around blog articles, which never influences scores or listings

Key features

• Free, independent directory of AI agent skills, continuously crawled and deduplicated from GitHub
• Every skill version pinned to a commit SHA, with validation, license checks, security scanning, and quality judging
• Sandboxed benchmarking that tests each skill with the same model, with and without the skill, publishing the Lift in percentage points
• Free MCP server that connects any agent to the skills you choose
• Per-agent loadouts: toggle skills and tools per agent so the agent only sees what was selected
• 0-100 trust score built from security, quality, license, maintenance, and community signals
• Upstream change tracking with re-scanning, re-benchmarking, and notifications
• Completely free to use — the directory, MCP connector, benchmarks, and alerts all cost nothing

Pros

• Only platform on this list focused on verifying and benchmarking agent skills rather than selling voice infrastructure
• Lift is always shown separately from ratings, and money never touches scores
• Security-first approach directly addresses the ~26% of skills found to carry vulnerabilities
• Per-agent loadouts give fine-grained control over exactly what each agent can see and use
• Everything is free — no per-minute fees, subscriptions, or paywalled features

Cons

• Currently pre-launch and actively being built — the waitlist/newsletter is live but the product is not yet public
• Not a voice framework itself; you still need a voice stack like Pipecat or LiveKit to build the agent

Visit SkillzBench

#2 Pipecat

Pipecat is a free, open-source Python framework from the Daily.co team built for engineers who want full control over how audio frames move through a voice agent. According to its documentation, the framework handles speech-to-text, voice activity detection, the LLM, and text-to-speech as composable stages, so you can assemble a pipeline that fits your exact latency, quality, and cost requirements. Integration coverage is notably wide, including Deepgram, ElevenLabs, Cartesia, Kokoro, Whisper, Gemini, and several dozen other providers. For teams comfortable with Python who want to own the audio pipeline in production, Pipecat is widely regarded as one of the strongest long-term foundations in the open-source voice space. The trade-off is that Pipecat ships nothing above the framework layer: there is no visual builder, no post-call analytics, and no CRM connectors. Every change to conversation logic means editing Python, committing, and redeploying, which is fine if you have engineering bandwidth and rough if you want a working system on day one.

Best for: Engineering teams comfortable with Python who want full control over their voice agent's audio pipeline

Pricing: Free (open source); Pipecat Cloud is available as a paid managed option — contact for cloud pricing

Key features

• Composable pipeline stages for STT, voice activity detection, LLM, and TTS
• Wide integration coverage including Deepgram, ElevenLabs, Cartesia, Kokoro, Whisper, and Gemini
• Bring-your-own-key support across every layer of the stack
• Fully self-hostable and free under open source
• Python-first framework suited to engineers who want to own the audio pipeline
• Pipecat Cloud available as a managed option to skip the ops side

Pros

• Maximum control over every stage of the audio pipeline
• Extensive provider integrations across STT, LLM, and TTS
• Free and self-hostable with no vendor lock-in
• Backed by the Daily.co team with an active community

Cons

• No visual builder, post-call analytics, or CRM connectors — everything above the framework layer is DIY
• Every conversation-logic change requires editing Python and redeploying
• Requires engineering bandwidth to build a production-ready system

Visit Pipecat

#3 LiveKit Agents

LiveKit Agents is a WebRTC-native, open-source voice framework from LiveKit, built on top of their widely used real-time media server. According to its documentation, it's organized as composable pieces: the core media server, the Agents framework for voice AI logic, and LiveKit SIP for PSTN bridging, so you can connect your agent to real phone networks. The underlying WebRTC infrastructure is battle-tested for demanding audio scenarios — packet loss, multi-party rooms, and browser-to-browser calls — which makes it a dependable foundation for teams building real-time voice experiences. A managed cloud option is also available if you don't want to run the media server yourself, handling scaling, geographic distribution, and SIP trunking for you. Multiple 2026 buyer's guides have named LiveKit Agents the leading open-source voice agent option, citing its Apache 2.0 license, self-hosting, and full customizability. Like Pipecat, it is a code-first SDK: there's no visual interface or built-in analytics, and getting a call out the door means wiring up the media server, the agent worker, and the SIP bridge separately.

Best for: Engineering and DevOps teams that need battle-tested WebRTC real-time infrastructure and full control over their voice stack

Pricing: Free (open source, Apache 2.0); managed cloud available — infra cost only when self-hosted

Key features

• WebRTC-native real-time media infrastructure built on LiveKit's widely used media server
• Composable architecture: core media server, Agents framework, and LiveKit SIP for PSTN bridging
• Apache 2.0 licensed and fully self-hostable
• Managed cloud option handling scaling, geographic distribution, and SIP trunking
• Battle-tested for packet loss, multi-party rooms, and browser-to-browser calls

Pros

• Proven real-time media infrastructure trusted for demanding audio scenarios
• Apache 2.0 license with no licensing cost and no vendor lock-in
• SIP support for connecting agents to the phone network
• Flexible choice between self-hosting and managed cloud

Cons

• Code-first SDK with no visual interface or built-in analytics
• Requires wiring the media server, agent worker, and SIP bridge separately
• Can be overkill for a standard inbound or outbound phone agent

Visit LiveKit Agents

#4 Dograh

Dograh is a free, open-source voice AI platform, licensed under BSD-2, that aims to bridge the gap between bare-bones frameworks and commercial platforms like Vapi, Retell, and Bland — while remaining self-hostable. According to its documentation, Dograh provides a visual workflow builder where you connect nodes instead of writing Python, so prompt tweaks don't require a redeploy, plus built-in telephony and post-call analytics out of the box. Production features like voicemail detection, call transfer, variable extraction, a knowledge base, and CRM connectors all come standard, with the same feature set whether you self-host or use the managed cloud. Dograh also supports BYOK (bring your own key) across every layer — Deepgram or Whisper for STT, ElevenLabs or Kokoro for TTS, and any LLM for the brain — and self-hosted models can be swapped in through the UI without code. It's a strong choice for teams that want open-source freedom but need to ship a production voice agent quickly rather than building the platform layer themselves.

Best for: Teams that want open-source self-hosting plus production features like a visual builder, telephony, and analytics without building the platform layer themselves

Pricing: Free (open source, BSD-2 license) with an optional managed cloud

Key features

• Visual workflow builder — connect nodes instead of writing Python, so prompt tweaks don't require redeployment
• Built-in telephony and post-call analytics included out of the box
• Production features standard: voicemail detection, call transfer, variable extraction, knowledge base, and CRM connectors
• BYOK across every layer: Deepgram or Whisper for STT, ElevenLabs or Kokoro for TTS, any LLM
• Self-hosted models swappable through the UI, no code required
• BSD-2 licensed and self-hostable, with an optional managed cloud

Pros

• Fastest path from open source to a production-ready voice agent on this list
• Visual builder means non-code changes and quicker iteration
• Production features (telephony, analytics, CRM connectors) included standard
• Same feature set whether self-hosted or on the managed cloud

Cons

• Newer project with a smaller track record than Pipecat or LiveKit
• Managed cloud pricing is not published — contact for pricing
• Less suited to engineers who want raw, code-level control over every audio frame

Visit Dograh

#5 EchoKit Server

EchoKit Server is an open-source AI service orchestrator focused on real-time voice use cases, written entirely in Rust for stability, safety, and speed. According to the project's documentation, it starts a WebSocket server that listens for streaming audio input and returns streaming audio responses, coordinating services such as LLMs, ASR, TTS, VAD, MCP, search, and knowledge/vector databases to generate intelligent voice responses from user prompts. It works with any OpenAI-compatible API provider for ASR — the project's tutorial demonstrates Groq's Whisper ASR service — and includes both device-side and server-side voice activity detection, with a Rust port of the popular Silero VAD. Notably for this list, EchoKit has first-class MCP support for actions, letting the voice agent call external tools during a conversation. The companion echokit_box project provides ESP32-based firmware for a hardware client, and a JavaScript web page can be run locally as a software client. It's a compelling pick for builders who want a lightweight, stream-everything orchestrator — including hardware voice devices — under their own control.

Best for: Developers building custom, fully self-controlled voice AI applications — including hardware voice devices — who want a lightweight Rust orchestrator

Pricing: Free (open source); users provide their own API keys for services like Groq and ElevenLabs

Key features

• Open-source voice AI orchestrator written entirely in Rust
• WebSocket server for streaming audio input and streaming audio responses
• Coordinates LLMs, ASR, TTS, VAD, MCP, search, and knowledge/vector databases
• Works with any OpenAI-compatible API provider (e.g., Groq's Whisper ASR)
• Device-side and server-side voice activity detection, including a Rust port of Silero VAD
• MCP support for configuring agent actions and tools
• ESP32-based hardware client (echokit_box) plus a local JavaScript web client

Pros

• Rust implementation prioritizes stability, safety, and speed
• Stream-everything architecture keeps multi-model orchestration efficient
• Built-in MCP support for connecting the voice agent to tools and actions
• Unique hardware option via the ESP32-based EchoKit device

Cons

• Requires comfort with CLI tools and running server applications on Linux-like systems
• ASR, LLM, and TTS require external API keys (e.g., Groq, ElevenLabs) unless self-hosted
• Smaller community and ecosystem compared to Pipecat or LiveKit

Visit EchoKit Server

Conclusion

Open-source voice AI has genuinely come of age in 2026. Whether you choose Pipecat for pipeline-level control, LiveKit Agents for battle-tested WebRTC infrastructure, Dograh for a fast path to production with a visual builder, or EchoKit Server for a Rust-based orchestrator with hardware support, you can now build voice agents that rival proprietary platforms — without per-minute fees or vendor lock-in. But going open source means assembling your own stack, and that stack increasingly includes agent skills and MCP tools pulled from the community — where published research found roughly 26% of skills carry at least one vulnerability. That's the gap SkillzBench is being built to close: a free, independent directory and free MCP connector that verifies every skill (pinned to a commit SHA, license-checked, security-scanned), benchmarks each one in a sandbox and publishes the Lift, and lets you equip each agent with only the skills and tools you've chosen. Everything — the directory, the benchmarks, the MCP server, the alerts — is free, and money never touches scores. Skills, proven. Join the SkillzBench waitlist today and be first in line when the platform launches.

Frequently asked questions

What makes SkillzBench different from the voice agent frameworks on this list?

Frameworks like Pipecat, LiveKit Agents, Dograh, and EchoKit give you the infrastructure to build a voice agent. SkillzBench solves a different problem: once your agent exists, which agent skills and MCP tools should it use, and are they safe? SkillzBench is being built as a free, independent directory that verifies skills (commit-SHA pinning, license checks, security scans, quality judging), benchmarks them in a sandbox and publishes the Lift in percentage points, and equips agents via a free MCP server with per-agent loadouts. A 0-100 trust score covers security, quality, license, maintenance, and community signals — and money never touches scores.

Is SkillzBench free to use?

Yes — everything on SkillzBench is free: the skills directory, the MCP connector, the benchmarks, and the alerts. SkillzBench earns revenue from advertising shown in and around its blog articles, never by charging users, and advertising money never influences scores or listings. Note that SkillzBench is currently pre-launch: the waitlist and newsletter are live, and the product is actively being built.

Why should I worry about security when using open-source agent skills?

Published research found that roughly 26% of agent skills carry at least one vulnerability. Open-source voice stacks often pull in community skills and MCP tools to extend what the agent can do — booking appointments, querying CRMs, routing calls — and each one is potential attack surface. SkillzBench's approach is to security-scan every skill, pin every version to a commit SHA, license-check it, and assign a trust score from 0-100 based on security, quality, license, maintenance, and community signals, so you know what you're equipping before your agent sees it.

Which open-source voice framework should I choose in 2026?

It depends on your team and use case. Choose Pipecat if you're Python-comfortable and want full control over the audio pipeline. Choose LiveKit Agents if you need battle-tested WebRTC infrastructure or multi-party real-time audio. Choose Dograh if you want open-source self-hosting plus a visual workflow builder, built-in telephony, and post-call analytics out of the box. Choose EchoKit Server if you want a lightweight Rust orchestrator or a hardware voice device. Whichever you pick, you'll be equipping the agent with skills and tools — which is where SkillzBench's free verification and benchmarking layer fits in.

What does 'Lift' mean in SkillzBench's benchmarks?

Lift is SkillzBench's benchmarking metric: every skill is tested in a sandbox with the same model, with and without the skill, and the resulting difference in performance is published as a Lift in percentage points. Importantly, Lift is always shown separately from trust ratings — a skill's measured performance gain and its security/quality score are never blended into a single number, and money never touches either.

Can open-source voice agents really compete with paid platforms like Retell or Vapi?

Increasingly, yes. Open-weight TTS models such as Kokoro, Chatterbox-Turbo, and Higgs Audio V2 now benchmark competitively against paid APIs, and frameworks like Pipecat and LiveKit let you wire in best-in-class STT, LLM, and TTS providers — or fully self-hosted models. The trade-offs are engineering effort and responsibility for security and compliance, since you own the whole stack. That's also why verification tooling matters: if you're pulling in community skills and MCP tools, use a free directory like SkillzBench to security-scan and benchmark them before your agent depends on them.

When does SkillzBench launch?

SkillzBench is currently pre-launch and actively being built. The waitlist and newsletter are live now, so joining gets you early access and updates as the directory, benchmarking sandbox, and free MCP connector roll out. In the meantime, the SkillzBench blog publishes guides on AI tools, agent skills, and automation workflows for business owners, developers, prompt engineers, and agent builders.