Repolode
Submit
Makson179

Makson179/Bello

A Python CLI that wraps coding agents like Codex and Claude Code with an optional supervisor, reviewer, adversary, and local log distiller, each on its own model.

View on GitHub
Stars
161
Forks
12
Open issues
3
Last commit
2 days ago

What is it?

Bello runs a coding task through a coder agent and four optional components: a runtime supervisor that watches the live run and judges risky commands, a completion reviewer that checks the result against the task, an adversary that tries to break the finished work with edge cases, and a local log distiller that shortens tool output before the coder reads it. Each role can use a Codex or Claude Code subscription or an API provider such as OpenAI, Anthropic, or OpenRouter, and the four components can be switched on or off independently.

The project targets developers who already pay for coding agent usage and want to spend it more efficiently. It also offers Smart Execution, which runs independent tool calls in parallel, an optional Goal mode that keeps a run going until the reviewer accepts the objectives, and No Need For Context, which summarizes older conversation history. The output of a run is written to a FINAL_REPORT.md with changed files, checks, and remaining risks. Bello supports macOS, Linux with bubblewrap, and native Windows 11 or Windows Server with a one-time sandbox setup.

The repository includes benchmark results on ProgramBench tasks comparing raw coder runs with Bello configurations, reporting cost, usage, and score differences. Installation is via pipx, and a plugin for Codex and Claude Code lets the agent itself run Bello and recommend a configuration.

Key features

  • 01Optional runtime supervisor that watches the run and judges risky commands before they execute.
  • 02Completion reviewer that starts with fresh context and checks the result against the task.
  • 03Adversary pass that probes finished work with edge cases and invalid input.
  • 04Local ModernBERT-based log distiller that shortens tool output without sending it to a paid model.
  • 05Each role can use a different model or provider, including Codex, Claude Code, OpenAI, Anthropic, and OpenRouter.
  • 06Smart Execution runs independent tool calls in parallel.

Use cases

  • Reducing Codex or Claude Code usage on long coding tasks while keeping output quality.
  • Mixing a cheaper coder model with a stronger reviewer or supervisor model.
  • Running coding tasks with a reviewer and adversarial testing before accepting the patch.
  • Running agent-driven coding in a sandboxed workspace on macOS, Linux, or Windows.

Quick start

  1. 1

    Install Bello with pipx

    pipx install bello
  2. 2

    Check what is ready

    bello doctor
  3. 3

    Log in to your Codex subscription (install the Codex CLI first)

    codex login

Things to know

  • Supports macOS, Linux with bubblewrap, and native Windows 11 or Server 2022/2025 only after a one-time sandbox setup.
  • Requires at least one model source such as a Codex or Claude Code subscription, or an API provider.
  • The log distiller model is about 599 MB and downloads on first run.
  • Some configurations take much longer to finish than a raw coder run.

Topics

#ai-coding#multi-agent#coding-agent#cli#llm#sandbox#code-review#agentic-ai#agentic-workflow#ai-agents

FAQ

What is Bello?
A Python CLI that wraps coding agents like Codex and Claude Code with an optional supervisor, reviewer, adversary, and local log distiller, each on its own model. Bello runs a coding task through a coder agent and four optional components: a runtime supervisor that watches the live run and judges risky commands, a completion reviewer that checks the result against the task, an adversary that tries to break the finished work with edge cases, and a local log distiller that shortens tool output before the coder reads it. Each role can use a Codex or Claude Code subscription or an API provider such as OpenAI, Anthropic, or OpenRouter, and the four components can be switched on or off independently.
Is Bello free to use? What license does it have?
Bello is open source under the MIT license. Check the license terms before using it in commercial products.
Is Bello still maintained?
Actively maintained. The last commit to Bello was on Oct 20, 2026.
What can I use Bello for?
Reducing Codex or Claude Code usage on long coding tasks while keeping output quality. Mixing a cheaper coder model with a stronger reviewer or supervisor model. Running coding tasks with a reviewer and adversarial testing before accepting the patch. Running agent-driven coding in a sandboxed workspace on macOS, Linux, or Windows.
How do I install Bello?
Install Bello with pipx: pipx install bello Check what is ready: bello doctor Log in to your Codex subscription (install the Codex CLI first): codex login
Where can I get Bello?
The source code and installation instructions are on GitHub: https://github.com/Makson179/Bello

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects