MahBot
Mahbot(i.e. my bot) is an agentic system that automates coding while also providing built-in tools for everyday research and media editing and generation.
Mahbot treats software work as a managed pipeline, not a chat session: you talk to the Manager about intent and scope; specialist agents do the actual work. Manager creates and refines tickets on the board, and only escalates real product decisions to you.
Reliability comes from orchestration and process, not based on the expectation that the current frontier model will one-shot any task.
Autonomy is achieved using the pipeline - you can request a large amount of work and manager will ensure that every piece is analyzed, implemented, reviewed, tested and commited.
Batteries included:
- Smooth native GUI for the core pipeline management as well as code editor, diff viewer and shell
- Modern agentic stuff: subagents, adversarial review and QA, deterministic diagnostics
- Telegram bot integration that allows you to easily manage the work from your smartphone
- Voice control using a CPU-optimized local speech-to-text model that turns babble into features (passive wake-word detection wip)
- Background maintenance process to clean up the usual videcoding bloat as well as other code quality issues
- Full history of the previous work in the tickets with efficient hybrid search over it
- Out-of-the-box workspace discovery for per-role contexts, auto-detected diagnostics commands. No need for plugins, AGENTS/CLAUDE/other.md files or custom configurations. Just add the API key and state your wishes
- General assistant adaptive agent for Q&A, research, prototyping, image & video generation/editing and basically any other purposes
OpenRouter is the default provider; manager-side roles default to GLM 5.3 Flash and worker-side roles to DeepSeek 4 Flash. A custom self-hosted OpenAI-compatible endpoint (llama.cpp, vLLM, or alike) can be configured in Settings for chat requests. Note that media tools are currently tied to OpenRouter — so its key is still needed for those even when a custom endpoint handles dev agents. Also, should work quite well with smaller models like Qwen 3.8 27b, and local + open-source mode is the primary long-term focus.
The Pipeline
Every ticket has a lifecycle with redundant checks:
| Phase | What happens |
|---|---|
| → Analysis | Parallel analysts research the ticket's assumptions & scope |
| → Planning | Manager sees the analysis and refines/cancels/approved or escalates |
| → Queued | Awaits in the engineer's queue according to it's priority |
| → Development | Engineer implements the ticket (or the required fixes) |
| → Diagnostics | Deterministic verification (format, lint, build, test) |
| → Review | Agentic verification focused on the code quality |
| → QA | Agentic verification focused on the product behaviour |
| → Sanitation | Audit untracked/new files in the working tree |
| → Done | Auto git commit with the ticket's title if the tree is dirty |
Circuit breaker pauses the work if a ticket goes through too many bounces, escalates to the manager and he handles from there.
Getting Started
Currently mahbot can only be installed from crates.io:
Then run mahbot to start the service, and you'll be asked to provide one of:
- OpenRouter API key, or
- a custom OpenAI-compatible endpoint
The rest of the setup will be explained and done through the admin Assistant, which comes with the mahbot_config tool. It will help you with configuring the workspace, adding other users, connecting the Telegram bot to comfortably direct mahbot agents from your phone using text, voice msgs & built-in commands, and setting up the search providers & chrome-use tooling for your agents.
Beware that as of now mahbot is only regularly tested on macos & linux, so it might still have unexpected bugs on other platforms.