Mesa Code

Mesa Code is an early public fork of Roo Code focused on one narrow goal: building a terminal-native coding agent that treats local AI runtimes as first-class systems, not just endpoint URLs.

The project is being shaped for developers who run local or self-hosted models and want the agent to know what is actually running: which model was requested, which model is active, whether it is ready, which runtime features are qualified, and how tool approvals behave across CLI, TUI, and automation surfaces.

Status

This fork is in active development and is not packaged as a stable release yet. The public roadmap is a draft and will change as the CLI architecture settles.

The product name is Mesa Code. The repository still contains forked Roo Code package names and compatibility paths while the rename is staged. The intended CLI command is mesa; roo will remain a compatibility alias during migration.

See ROADMAP.md for the current public plan.

Direction

• CLI and TUI first
• local/self-hosted runtimes first
• no required cloud account in the happy path
• OpenAI-compatible and Anthropic-compatible local endpoint support
• runtime readiness and model identity checks before claiming success
• first-class support for local runtime observability
• stable text, JSON, and stream contracts for automation
• explicit user control over tools, approvals, model selection, and config
• editor integration later, after the CLI core is solid

Local Runtime Goals

Mesa Code is being shaped around local model engines such as:

• vllm-mlx
• llama.cpp
• OpenAI-compatible local servers
• Anthropic-compatible local servers

The CLI should not hide local runtime complexity behind fake green buttons. If a runtime feature is not qualified, the CLI should say so clearly and fail closed.

Why Not Just Another Coding CLI?

Pi is impressively light. OpenCode and Kilo Code are strong agent CLIs. Mesa Code is aimed at a different gap: the local runtime/operator layer.

The goal is for the CLI to understand:

• what model was requested
• what model is actually serving
• whether the requested model is ready
• whether runtime features are qualified for that model
• what the runtime health and queue state look like
• whether approval behavior is consistent across TUI, CLI, and stream modes

That makes local inference easier to trust and easier to automate.

Development

Prerequisites:

• Node.js 20.x
• pnpm

Install dependencies:

pnpm install

Build the CLI:

pnpm --filter @roo-code/cli build

Run CLI checks:

pnpm --filter @roo-code/cli check-types
pnpm --filter @roo-code/cli test

Run monorepo checks:

pnpm check-types
pnpm lint

Repository Layout

• apps/cli - terminal CLI and TUI work
• src - existing Roo extension/runtime code being carved apart
• packages/core - shared core logic
• packages/types - shared contracts and provider/model types
• webview-ui - upstream webview UI retained while the fork narrows
• ROADMAP.md - public draft roadmap

Contributing

This fork is not ready for broad drive-by contribution yet, but focused collaboration is welcome around:

• CLI/TUI session architecture
• local runtime adapters
• terminal UX
• model/runtime observability
• privacy-first defaults
• tests for command-line and PTY behavior

Open an issue or discussion before starting large changes so the work lines up with the fork direction.

License

This repository remains under the upstream Apache 2.0 license.