Fixie AI

Fixie AI is a developer-focused platform for building voice AI agents that handle phone calls. The core product lets developers deploy agents that can take inbound calls, make outbound calls, access APIs mid-conversation, and complete tasks without human involvement. Fixie was founded by former Google and Apple engineers in Seattle and focuses on the technical infrastructure that makes voice agents reliable at production scale: low-latency speech recognition, natural turn-taking, and function calling during live conversations. Pricing is usage-based. The platform competes with Vapi, Retell AI, and similar voice agent infrastructure providers.

Fixie was founded in 2022 by Matt Welsh, former Google engineering director and Harvard CS professor, alongside other engineers with backgrounds in speech AI and ML infrastructure. The founding team is notably technical, which shows in the product's emphasis on the underlying speech pipeline rather than on no-code visual builders.

The product launched in 2023 as a voice AI platform for developers. The timing was good: the combination of improved large language models and better real-time speech processing made functional AI phone agents viable for the first time. Fixie's bet was that there would be significant demand for the infrastructure layer, not just turn-key products.

Why voice agents are hard

Building a voice agent that sounds natural and works reliably is a different engineering problem than building a text chatbot. The latency requirements are different. In text, a user can wait a second or two for a response. On the phone, anything over about 800ms starts to feel wrong. Callers interpret long pauses as dropped calls or system failures.

This means the entire pipeline from speech recognition through response generation through text-to-speech has to be optimized for real-time performance. Streaming ASR processes audio as it arrives. LLM generation has to start before transcription is complete, using predictive buffering. TTS output starts streaming before the full response text is ready. Each of these optimizations reduces latency by 100-200ms; combined they make the difference between a conversation that feels natural and one that feels like talking to an old IVR system.

Fixie's team has spent real engineering effort on this pipeline. The result is voice conversations with latency that is noticeably better than simpler implementations, though it is not zero and it degrades when function calls require slow external APIs.

Function calling during live calls

The feature that makes Fixie's voice agents genuinely useful rather than just able to chat is function calling. During a live phone conversation, the agent can invoke code, query APIs, and use the results in its response, all in a time window that does not break the conversational flow.

The architecture for this is: you define functions in your Fixie agent configuration, the same way you define tools in an OpenAI or Anthropic API call. The LLM decides when to invoke a function based on the conversation, calls it, gets the result, and incorporates the result into the next response. For a voice agent, this all has to happen within roughly 300-400ms to stay inside acceptable latency budget.

The practical applications are significant. An appointment scheduling agent can check real availability, not just say "let me check and call you back." A support agent can look up account status. An ordering agent can confirm real inventory levels. These are the capabilities that turn a voice demo into a production deployment that handles real customer interactions.

Telephony integration

Fixie handles the AI layer; the telephony connection comes through integration with providers like Twilio, Vonage, or via SIP for direct carrier connections. The web dashboard lets you configure phone numbers, route incoming calls to agents, and set up outbound call queues.

For most developers building production voice agents, the Twilio integration is the most common path. You configure a Twilio phone number to route to Fixie via webhook, and Fixie handles the conversation from there. The setup is documented and works reliably once configured.

SIP integration is available for organizations with direct carrier connections or existing PBX infrastructure that they want to route through Fixie without going through a third-party telephony provider.

The SDK experience

Fixie provides SDKs for JavaScript and Python, with a REST API for other languages. The JavaScript SDK is more mature. The Python SDK covers the core use cases but has some gaps in the more advanced features.

A basic inbound call handler is straightforward to implement: you define an agent with a system prompt, configure some functions, connect a phone number, and deploy. The complexity scales with the sophistication of what the agent needs to do.

The documentation is the area where Fixie has the most room to improve. Core functionality is well-documented with good examples. Advanced configurations, particularly around custom TTS providers, multi-party call handling, and complex function call chains, are less thoroughly documented. The GitHub repository has example projects that fill some of these gaps.

Pricing reality

Usage-based pricing means the cost scales with actual usage, which is good for unpredictable workloads but harder to budget for. A free tier for development and testing is available and sufficient for building and testing an agent before production deployment.

Production pricing is per minute of conversation. The actual rate varies and Fixie adjusts it with model and infrastructure costs, so the current rates are best checked in their pricing page or by signing up. For planning purposes: voice AI platforms in this space typically run between $0.05 and $0.20 per minute of conversation, with volume discounts at scale. At those rates, a contact center handling 10,000 minutes of AI conversations per month might pay $500-$2,000 per month for the AI conversation infrastructure, excluding telephony costs.

Enterprise pricing is custom and adds volume discounts, SLA commitments, dedicated support, and options for private deployment.

Where Fixie fits in the market

The voice agent infrastructure market has several credible players: Vapi, Retell AI, Bland AI, and Fixie are the names that come up most often among developers building voice agents in 2026. Each has slightly different strengths.

Fixie's differentiation is its engineering depth on the speech pipeline and function calling reliability. Vapi has a larger community and more third-party integrations. Retell AI has strong documentation and a more polished no-code configuration experience. Bland AI targets the outbound call use case specifically.

The choice between them is less about which is objectively better and more about which fits your specific technical context. Testing the latency and function calling reliability on your actual use case, with your actual external APIs, on your actual infrastructure, is the right evaluation approach. All of these platforms have free tiers that let you do that testing before committing to a production deployment.

Key features

• Real-time voice agents that handle phone calls with sub-second latency
• Inbound and outbound call handling with programmable logic
• Function calling so agents can query APIs, look up data, and take actions mid-call
• Webhook integration for real-time event handling during conversations
• Custom voice selection with multiple TTS providers
• Telephony integration via SIP or direct PSTN through Twilio and other providers
• SDKs for JavaScript, Python, and REST API

Pros and cons

Who is Fixie AI for?

• Appointment scheduling bots that handle inbound calls for healthcare or service businesses
• Outbound call campaigns for lead qualification or survey collection
• Customer service deflection for common inbound inquiry types

Alternatives to Fixie AI

If Fixie AI isn't quite the right fit, the closest alternatives are synthflow , air-ai , decagon-ai , hume-ai , vapi , and retell-ai . See our full Fixie AI alternatives page for side-by-side comparisons.