pedro castanheiraPT
← All work

Personal project · open source

Cast Code

Built a multi-agent coding CLI that plans, delegates to specialists, uses tools over MCP and checks its own work in the terminal.

Role
Product, architecture and development
Status
Published on npm · evolving
Stack
  • TypeScript
  • Node.js
  • NestJS
  • MCP
  • WebSocket
you in theterminalprompt + @file, @git, URLcodebase +project skillsmain agent(plan)model per role:planner, coder, reviewer…specialists(coder, architect…)tools + MCP(~30 templates)/remote: web + voiceverification:review + testsself-critiquecommit/ PRpublished on npm: cast-code

Problem

Linear assistants write code fast but lose conventions, do not verify the result and break the flow between plan, implementation and delivery.

Available data

The local codebase, the Git diff, files and URLs mentioned in the prompt, and the skills and agents defined in the project itself.

Architecture decisions

A model per responsibility

Planner, architect, coder and reviewer can use different providers and models, balancing cost and quality per step.

Extensible through files

Skills and agents live in the project as Markdown and are versioned with the code — no changes to the CLI core.

An explicit loop

Plan, act, verify and self-critique before delivering a commit or PR.

Implementation

  • A main agent that delegates to specialists defined in Markdown (coder, architect, reviewer, tester, frontend, backend, devops and others).
  • A live tree of subagents in the terminal, with status, current tool and time.
  • AI Git commands: /up (commit), /split-up (splits commits by intent), /pr and /review.
  • An MCP client with a template catalog, /remote to follow along in the browser (with voice) and /bridge to run through other providers’ CLIs.

Result

  • npmpublished as cast-code

Published on npm and continuously evolving.

Limits and lessons

  • Personal project; there are no public usage metrics.
Next caseCast Skills →