Quick start · What it does · Documents · Terminal UI · Roadmap
A plugin-first agent harness in Rust. An agent that writes code, runs tools and keeps a memory of why the project is the way it is, where almost every part is an interface you can replace.
Quick start
git clone https://github.com/codefather-labs/harness-multitool
cd harness-multitool
./run.sh # builds it and opens the terminal UI (Windows: run.ps1)
Open the settings with Ctrl+, and choose a provider: a key for an API, a local model on Ollama, or your ChatGPT plan through gpt-subscription, which signs in on OpenAI's own page.
What it does today
- The agent loop in a terminal UI: streaming answers with Markdown and tables, tool calls with permission prompts, several tabs, sessions that resume, and what the agent is doing on the status line.
- Models: Anthropic and OpenAI-compatible APIs, Ollama and other local servers, a ChatGPT plan. The model list comes from each provider, with prices where they are known.
- Plan mode: the agent reads and plans, and nothing in the project can change until you approve.
- Step-debug mode: the agent walks a running pipeline step by step through a debug router built by one contract, and every finished cycle is saved.
- Memory: a project journal, mind maps and live flowcharts, which the agent's context is built from.
- Done means proven: a task finishes only when its checks pass, and the evidence goes into the journal.
- Callbacks and channels: events from outside reach the agent as data; conditions, custom handlers, and a Telegram channel.
- Hot reload: rebuild and swap the binary in the same terminal, with the sessions going on.
What it is meant to be
- Plugin-first. The core is an agent loop and an event bus. Model providers, authentication, tools, permissions, context, memory, sessions and routing are interfaces, and the built-in ones are plugins of those same interfaces.
- Modes are frameworks, not prompts. A mode gives the agent its rules and tools, and the harness holds to them.
- Many agents, many machines. Sessions become addressable, message each other, and spread across machines and models.
- Runs in the browser. The core compiles to WebAssembly; the project site is to host the CLI itself as a live demo.
Documents
| Principles | what the project optimises for, and what it refuses |
| Invariants | what must hold at all times, and how each is checked |
| Constraints | rules, limits and constants |
| Interfaces | every pluggable component: its contract and its boundaries |
| Architecture | how the pieces fit, with a map |
| Terminal UI | running it, keys, debug commands, screens as images |
| Agent workflow | what an agent must do in each situation; ships with the harness, and agents are given its path |
| Tools and modes | what an agent can call, and how the harness's modes differ from utilities |
| Callbacks | how outside apps reach an agent, and how to write a project's own callback handler in Rust: the contract |
| Design | the logo, icons, themes and the mockups the UI is built against |
| Decisions | why the project is the way it is (ADRs) |
| Roadmap | the order of work |
| Open questions | what is not decided yet |
| References | projects we learn from |
Every rule in these documents has an id (P-…, INV-…, CON-…, IF-…). Code, tests and decisions cite them.
On authentication and subscriptions
The project ships authentication for API keys, and a subscription sign-in only where the vendor documents one for apps like this (ADR 0019). A ChatGPT plan is signed in through OpenAI's flow for open-source apps. A Claude subscription is not offered: Anthropic's terms do not permit it in other apps, so Anthropic models take an API key.