Skip to content
davidpp.com
tokebak

An agent harness for voice and text, written in Rust.

A speech model talks with you. When it needs a fact, a backend agent looks it up and answers with its sources. Longer jobs go to other agents. I want it fast, and I want to see where every fact came from.

Research harness. Two of my other projects run on it. Sources

A support call, step by step

Each step shows what the caller hears, what the backend is doing, and the events Tokebak logs for it. From step 3 on, you can cut in as the caller.

  1. Question

    CallerCan I still return the hiking boots from order 2187?

    AgentLet me check.

    hand-off
    Backend

    Gets the request. The speech model hands off anything that needs a fact, and it is told never to guess while it waits.

    • voice.delegation
  2. Policy search
  3. Early update
  4. Order lookup
  5. Answer

Step 1 of 5: Question. Caller: Can I still return the hiking boots from order 2187? Agent: Let me check. Backend: Gets the request. The speech model hands off anything that needs a fact, and it is told never to guess while it waits.

How it’s built

One Rust process. Voice goes from the browser straight to OpenAI’s Live API, never through the server. Typing and talking share the same conversation, so you can switch mid-task.

Agents
A folder with AGENT.md, tools and skills. A conversation keeps the version it started with.
Tools
JavaScript in QuickJS, or MCP servers. A retried write doesn’t run twice.
Knowledge
SQLite search per agent, indexed ahead of time, never during a call.
Citations
Answers cite passages and tool results. A citation that doesn’t resolve says so.
Trace
Every request, search, tool call and voice event, with cost, exported as OpenTelemetry.
Longer work
Goes to other agents over ACP or A2A, or to local subagents. When one needs a decision, it asks you through Tokebak.

What runs on it

icitte, my voice and canvas workspace, and our family assistant in FamilyOS. The sample stores in the repo run on made-up customers and orders, to give it realistic work.

What’s next

Checking each answer against its sources while the conversation keeps going, so a wrong fact gets caught without making you wait. Then reaching you with a result when no session is open.

Sources

Last checked .

Status
Research harness
  • Tokebak READMEREADME at b6cfe8f (private)

    What it is, the voice and backend split, agent folders, JavaScript tools, SQLite search, citations, the trace, handing longer work to ACP agents, A2A workers and local subagents, and the sample stores' made-up customers and orders.

  • Direction, contracts and voice runtimedocs/direction.md, docs/contracts.md, ADRs 0039 and 0049, and src/runtime/live.rs at b6cfe8f (private)

    The goal and its three priorities (speed, trust, continuity), the hand-off, the cited early update, discarding an answer when the caller speaks again, asking the person on behalf of another agent, what isn't built yet, and the event names in the demo.

  • icitte and FamilyOSicitte README at c856fb9; FamilyOS infra/tokebak at 6bf9a5c (both private)

    Both run on a pinned Tokebak commit.