tlbx 10.22: Voice Control for Coding Agents, Linux Downloads First

tlbx 10.22.0, released on October 19, brings direct voice control of coding-agent sessions into its stable channel. You can use voice to inspect existing work, select a session and send an agent its next task. This release supplies Linux arm64 and x64 downloads; Windows and macOS builds are explicitly unavailable.

The stable release notes collect changes since 10.19.3, including voice orchestration first introduced in the 10.21.0 development release. The useful change is a way to manage agents already running on your machine from the same browser workspace.

What tlbx does with your coding agents

Previously named MidTerm, tlbx is a self-hosted terminal browser multiplexer. Install it on the workstation or server where your repositories and tools live, then connect through a browser on a desktop, tablet or phone.

Its versioned guide describes two session types. A Terminal Session displays an ordinary shell or command-line application. An Agent Controller Session presents a supported coding agent as a conversation, with tool calls, code changes, questions and approval controls.

Both use the same workspace for session navigation, files, Git changes and app previews. Built-in agent launch options include Codex, Claude Code, Grok Build and OpenCode, with features depending on the agent and its configured provider. The native OpenCode connection requires CLI 1.14.19 or newer.

This is useful when work is spread across several sessions. A developer might have one agent implementing a change, another reviewing it and a terminal running tests. Keeping those sessions visible together can make the next instruction easier to place, provided the developer still checks which session is active.

What voice can actually operate

The release notes describe GPT-Live 1 running directly inside tlbx to inspect, create, select and steer existing terminal and Agent Controller sessions. The tagged voice-tool catalog gives that description a specific scope.

Voice can list sessions and read their activity, select one in the initiating browser, create an ordinary terminal session and send a task through an existing agent's prompt API. It can also read bounded terminal scrollback. Agent Controller activity comes from its runtime state rather than a scraped terminal display.

The catalog instructs the prompt tool to avoid interrupting existing work unless requested. Its raw terminal-input and session-closing tools explicitly require browser confirmation. These definitions are useful when assessing the intended behavior, although they do not establish that every possible agent action is safely handled.

For a first trial, ask voice to describe the current sessions and select the one you name. Then send a small task whose target is easy to verify. Check the visible session and resulting prompt before trying commands that interrupt work or affect files. A misheard instruction becomes more consequential when it can reach a running shell.

The interface also gains compact, expandable voice-tool activity with running, completed, failed and declined states. Keeping those results attached to the original calls helps users see what happened when speech messages arrive between a request and its response.

Linux availability takes precedence over the platform badges

The general tlbx website and README describe Windows, macOS and Linux support. The 10.22.0 release page is more specific: only Linux arm64 and Linux x64 archives are available for this version. It lists macOS arm64/x64 and Windows x64/x86 as unavailable.

That distinction matters for someone updating an existing installation. A cross-platform project can publish a particular release for fewer targets. Check the release assets for the machine that will host the agents before planning an upgrade around the voice feature.

The native installer sets up password-protected HTTPS and offers service or user mode. The separate npm trial launcher is published independently and may lag native releases, according to the pinned guide. Neither route supplies the coding agents' model subscriptions for you.

Costs, data access and session lifetime

tlbx is free software under GNU AGPL v3, with commercial licensing also available. Your agent provider's usual charges still apply. Voice settings include provider API credentials and an inference-worker choice; the release does not advertise a bundled allowance of free voice or model usage.

The host keeps the repositories and running processes. Coding agents may still send relevant data to their model providers under their configuration and terms. Teams should review those connections before exposing private project context to a voice-controlled session.

The release notes describe replacing an external voice server and browser-delivered shared password with authenticated local voice and dictation endpoints. This is a documented implementation change, not an independent security assessment. The project's remote-access guidance recommends a trusted LAN or private VPN, with HTTPS and password authentication enabled.

Closing the browser leaves sessions running while the host stays awake and online. Shutting down the host stops its processes. Voice access from a phone therefore depends on the same machine and network availability as any other remote session.

For developers already using several coding agents, 10.22 gives voice a concrete role in checking and redirecting that work. Start by verifying session selection and visible tool results. Those small interactions will show whether speaking saves effort in your setup before you rely on it for a longer coding task.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help developer-tool teams explain new workflows, deployment requirements and product tradeoffs through technical stories that reach the right users.

Book a 30-minute call.