About
Framework
The architecture behind Codmir — how the CLI, daemon, SDK, and AI agents work together to give you a full development team on every project.
Overview
Codmir is an AI-native workspace where voice, code, and agents are unified into a single development pipeline. Instead of switching between a dozen tools, you install one CLI, add one SDK, and get an always-on AI team that reads your codebase, writes code, runs tests, and opens PRs — all behind a review gate you control.
Architecture
The platform is built on three layers that communicate through typed schemas:
1. CLI + Daemon
The codmir CLI is your primary entry point. It installs alongside a background daemon that keeps your project connected to AI agents around the clock — even when your terminal is closed. The daemon manages agent processes, heartbeats to the platform, and provides a local execution target for tickets.
2. SDK
@codmir/sdk instruments your application with error tracking, session replay, and agent hooks in a single import. It connects your running app to the platform so agents can observe production behavior, detect regressions, and act on real data.
3. Platform
The Codmir platform (web + API) handles authentication, project management, ticket workflows, and the review gate. It coordinates between your local daemon and cloud infrastructure, routing work to wherever you choose.
The Agent Loop
Every piece of work in Codmir follows a closed loop: observe → file → work → verify → review → remember.
- Observe — The SDK and Overseer agent monitor your application for errors, performance regressions, and user behavior.
- File — Issues are automatically filed as tickets with full context — stack traces, session replays, and suggested fixes.
- Work — The Coder agent picks up tickets, reads your codebase, writes implementation code, and runs your test suite.
- Verify — Tests pass, type checks succeed, and the agent validates the fix against the original error.
- Review — Every change lands in your inbox with a full diff, test results, and the agent's reasoning. You approve, request changes, or reject.
- Remember — Decisions, patterns, and fixes are persisted in project memory. Agents get smarter the longer they work with your code.
Execution Targets
You decide where work runs, per ticket:
- Self-hosted — Run on your own machine via the local daemon. Fast, private, and free. The daemon manages agent processes and reports results back to the platform.
- Cloud — Scale to cloud infrastructure when you need it. Auto-scaling compute with usage-based billing. Same review gate, same agent loop.
Review Gate
Nothing touches your codebase without passing through the review gate. Every AI-generated PR shows a full diff, test results, and the agent's reasoning. You can set the review mode per project:
- Manual — Every change requires human approval.
- Semi-auto — Low-risk changes (formatting, deps) auto-approve; everything else requires review.
- Auto — AI reviews on your behalf within an SLA. Escalates to human when confidence is low.
Voice + Chat
Give instructions in natural language — voice or text. The agent reads your codebase, writes code, runs tests, and iterates until the task is done. Voice mode uses a direct audio stream (no browser TTS), so you get real-time conversation with your project.
Memory
Every code review, architecture decision, bug fix, and user preference is stored in project memory. When an agent tackles a new task, it draws on everything it has learned about your codebase — naming conventions, preferred libraries, past decisions and their reasoning. The longer it works with you, the less you need to explain.
Get Started
Ready to set up your project? Follow the Getting Started guide to install the CLI, add the SDK, and activate your first agents.