KiraBridge: From a spoken idea to local coding agents

#agents #ai #llm #softwaredevelopment

I want to develop software while discussing an idea naturally — even on my phone while I’m out. That’s why I’m building KiraBridge and several companion tools.

The guiding principle is that we retain control: We decide which tasks the AI carries out, which credentials and services it may access, and which limits apply. Decisions and results should remain traceable.

Another goal is to stop losing project context and having to constantly search for earlier agreements, files or work states. Tasks, decisions and results should stay with their respective project so that individuals, and later teams too, can pick up their work where they left off.

The workflow starts with a voice conversation with ChatGPT. We discuss an idea, and I explicitly authorize its implementation. KiraBridge passes this task to Codex or Grok, running as local command-line tools on my computer. Questions, progress updates and results can then return to the conversation.

For example, I describe a change on my phone, discuss the details and request its implementation on my computer. Not every companion tool needs to participate.

The entry point should not remain limited to ChatGPT: Grok and other assistants with natural-language interaction should also be connectable through suitable interfaces. The workflow described and tested so far uses ChatGPT.

ThreadDesk is the shared working memory: It brings together threads, tasks, decisions and project knowledge. We currently use it individually; teamwork is planned so that people in smaller teams and agencies can later work together on projects and pick up where they left off.

4AllPass is a password manager with multiple locally stored vaults. It can also manage API keys; scoped team sharing and protected access for AI are planned extensions.

TollGate provides cost and usage control before execution: It accepts only connected calls and limits their usage according to configured rules and budgets. We do not promise a general cost cap for arbitrary signed-in CLIs; what matters is which execution paths are securely connected.

Taken together, ThreadDesk remembers the work, 4AllPass limits access and TollGate limits the usage routed through it.

Agent Authority examines available evidence of agent activity and highlights gaps or contradictions.

Gnom-Hub coordinates multiple agents.

BlockDesigner supports visual design; handing these designs over for implementation is a planned next step.

These tools are independent and can be combined as needed.

We distinguish between API access with keys, signed-in command-line tools and local models. MCP carries tool requests and results; it does not replace authentication or a provider subscription.

Individual real Codex and Grok runs have been documented. KiraBridge, ThreadDesk and TollGate have MCP servers in their source code. This does not mean that every connection has been verified together in operation. Protected access and parts of the overall integration are still under development.

I’m interested in your experiences: Would you use a voice-driven workflow like this? What feedback, controls and evidence would you expect before using it in your own projects?

This is an independent project developed with AI assistance, with no affiliation with OpenAI.

Links

Website: https://netzwerkpunkt.de/
GitHub profile: https://github.com/landjunge

Public repositories: