Here's a test we'd run before signing for any voice agent stack. Pull up the last five changes your phone flows needed, say a renamed CRM field that broke a lookup, or a model your agent relied on getting retired. Then ask, for each option on your shortlist, who actually makes that change.
Run that on Synthflow and LiveKit and you get two very different answers. That tells you more than a feature grid will, because launch day is the easy part. The agent spends years after it being changed.
On Synthflow, whoever owns the builder makes the fix
Synthflow describes its product as a visual flow builder plus APIs. Where the agreement includes launch support, its forward-deployed engineers implement, test and launch the agents with your team. After that, an operations lead opens the flow and edits a prompt or a routing rule. Anything the builder can't express goes through Synthflow's APIs or a support request under the terms of your agreement.
The model question is where it gets interesting. Model usage is part of Synthflow's agreement, so when the model behind your agent is retired, the swap is Synthflow's call. For a contact center without developers, that's a relief. For a team with strong opinions about models, it's a constraint to know about before signing.
The commercial shape follows the same logic. Synthflow sells one annual enterprise agreement priced against the call volume you expect, with no per-minute rate and no free trial. Transcription, voices, CRM connections and SMS sit inside it alongside the models. Your flows and prompts live in Synthflow's platform and the code is closed, so the natural moment to change course is the renewal.
It also plugs into a lot without custom work. Synthflow lists built-in connections to Zoom, Five9, RingCentral, 8x8, Twilio, Telnyx, Vonage and Freshcaller, with Genesys Cloud through an Enterprise setup.
On LiveKit, every fix is a pull request
LiveKit Agents is Apache-2.0, and it gives engineers full control over turn-taking, tools and models. Nothing about the agent is out of reach. The flip side is that every change is a code change. A new transfer rule goes through review and tests, then ships through LiveKit Cloud or your own servers.
LiveKit does have a browser Agent Builder, and LiveKit is upfront that it's meant for prototypes. It doesn't support workflows with handoffs or tasks, model plugins or tests. When an agent outgrows it, "Download code" turns it into a Python project, and from that point the agent belongs to your engineers.
So back to the five changes. On LiveKit every one of them lands on an engineer's desk, next to the dependency upgrades any codebase needs anyway. So does the product layer around the agent, from calling lists to QA on live calls, which nobody ships for you.
The bill reflects that split. LiveKit Cloud starts Build at $0 with 1,000 agent minutes, Ship at $50 a month with 5,000 and Scale at $500 with 50,000, then $0.01 per agent minute. Models and phone minutes come on top. The line no plan lists is engineering time, and on a code framework it's usually the biggest one.
One licensing detail matters for anyone with a strict open-source policy. The turn detector ships under a separate LiveKit Model License, so read it before it goes into production.
Where the calls run
Synthflow runs only in its own cloud. You pick a US or EU workspace at setup, and Synthflow notes that connected providers can process some requests elsewhere. Its Trust Center lists HIPAA as self-attested, which a healthcare buyer should read closely.
LiveKit gives you more places to run. A LiveKit Cloud project has a data region, with an EU option chosen at creation, and LiveKit lists a signed BAA from its Scale plan upward. The open framework and media server run in any data center you like. Running them yourself keeps call audio inside your network only when the models run there too, since every provider on your keys still receives the audio it works on.
The gap in the middle
Some teams want LiveKit's open code without turning every prompt tweak into an engineering ticket. That's the gap we built Dograh, the open-source AI voice agent platform, to sit in. Later changes are typed in plain words in the editor's Builder tab. Edits save as a draft, and the published version keeps taking calls until someone publishes the new one, while engineers can still read every line under the BSD 2-Clause license.
Self-hosted, it has no per-minute platform fee. Servers, model keys and phone minutes still cost money. Dograh Cloud charges a platform fee of 1 cent a minute.
For the maintenance years specifically, the QA node scores every step of test and live calls at a sample rate you set, and View Trace shows the speech-to-text, AI and voice time on each turn, which is usually where a "the agent feels slow" complaint gets settled.
Whatever you pick, write down who owns each kind of change before you sign, and test the first version over a real phone call. A browser demo won't show you what callers hear.
Which of your last five changes would have been hardest to make on the stack you run today?
Billing, data location and what each option ties you to are set out side by side in our Synthflow and LiveKit comparison for teams planning year two.
Top comments (0)