@magnitudedev/cli 0.0.1-alpha.3 → 0.0.1-alpha.4

package/README.md

@@ -2,19 +2,13 @@
   <img src="wordmark.svg" alt="Magnitude" width="100%" />
 </p>
 
-<p align="center"><b>Open source AI coding agent</b></p>
+<p align="center"><b>Open source coding agent</b></p>
 
 <p align="center">
   <a href="https://docs.magnitude.dev" target="_blank"><img src="https://img.shields.io/badge/📕-Docs-232f41?style=flat-square&labelColor=232f41&color=gray" alt="Documentation" /></a> <img src="https://img.shields.io/badge/License-Apache%202.0-232f41?style=flat-square&labelColor=232f41&color=gray" alt="License" /> <a href="https://discord.gg/VcdpMh9tTy" target="_blank"><img src="https://img.shields.io/discord/1305570963206836295?style=flat-square&logo=discord&logoColor=white&label=Discord&labelColor=5865F2&color=gray" alt="Discord" /></a> <a href="https://x.com/usemagnitude" target="_blank"><img src="https://img.shields.io/badge/-Follow%20Magnitude-000000?style=flat-square&labelColor=000000&color=gray&logo=x&logoColor=white" alt="Follow Magnitude" /></a>
 </p>
 
-Most coding agents are one long conversation that compacts and loses context. Magnitude is a multi-agent system that stays sharp over long horizon tasks.
-
-- **Advanced orchestration** — Orchestrator delegates to subagents with two-way communication
-- **Context stays clean** — Conversation stays focused on your intent. Less compaction, better UX
-- **Parallel by default** — Spins up multiple subagents working concurrently, not one thing at a time
-- **Native browser agent** — Vision-based browser agent verifies UI changes and behavior
-- **Any model** — Frontier, open source, local, whatever you want
+Open source coding agent where everything happens in specialized subagents. We call this *subagent driven development*. The main agent stays focused on your intent, leading to stronger performance on longer tasks and higher quality code.
 
 <p align="center">
   <img src="interface.png" alt="Magnitude Interface" width="100%" />
@@ -24,12 +18,9 @@ Most coding agents are one long conversation that compacts and loses context. Ma
 
 ```bash
 npm install -g @magnitudedev/cli
-bun add -g @magnitudedev/cli
-pnpm add -g @magnitudedev/cli
-yarn global add @magnitudedev/cli
 ```
 
-After installation, just run `magnitude` in your terminal to launch the TUI:
+After installation, navigate to the directory you want to work in and launch the TUI:
 
 ```bash
 magnitude
@@ -37,48 +28,61 @@ magnitude
 
 This will launch Magnitude with a setup wizard for configuring providers and models.
 
-Optionally, run `/init` to automatically set up an [AGENTS.md](https://agents.md) for your project.
+### Providers
 
-### Skills
+Magnitude works with most major model providers out of the box.
 
-Magnitude supports the [Agent Skills](https://agentskills.io) standard. Skills are automatically discovered on launch from the following locations in order of priority:
+**Subscription-based** — Use models you already pay for, no API key needed.
+- **ChatGPT Plus/Pro** — Sign in with your OpenAI account
+- **GitHub Copilot** — Sign in with your GitHub account
 
-1. `<project>/.magnitude/skills/`
-2. `<project>/.agents/skills/`
-3. `~/.magnitude/skills/`
-4. `~/.agents/skills/`
+**API key** — Bring your own key from any of these providers.
+- Anthropic, OpenAI, Google Gemini, OpenRouter, Vercel AI Gateway, Google Vertex AI, Amazon Bedrock, MiniMax, Z.AI, Cerebras
 
-## Features
+**Local models** — Connect to Ollama or any OpenAI-compatible endpoint. No auth required.
 
-### Orchestrator → subagent architecture
+See the [providers docs](https://docs.magnitude.dev/configuration/providers) for setup instructions. If you would like another provider to be supported, please feel free to create an issue or raise a PR yourself.
 
-The orchestrator manages the conversation and delegates to specialized subagents (builder, debugger, explorer, planner, reviewer, browser), each with its own context window, toolset, and permissions. Subagents do focused work and report back, keeping the orchestrator's context clean.
+## Why Magnitude?
 
-### Context sharing via artifacts
+- **Avoids the dumb zone.** Since building happens in subagents, the main agent stays focused on your intent and effective even on longer, messier tasks.
+- **Steerable subagents.** Subagents can surface blockers, ask clarifying questions, and be paused, resumed, or redirected mid-task.
+- **Throw out the skills files.** Plan, review, debug, and browse are built-in subagents. No configuration needed, and the main agent actually uses them consistently.
+- **Mix your models.** Chat with Claude, build with Codex, or use whatever combination fits your preferences for intelligence, speed, and cost.
+- **Scoped toolsets, not modes.** Real coding jumps between planning and building too often for explicit mode toggles. Each subagent gets tools scoped to its role instead.
+- **Talk to it while it works.** The main agent isn't blocked by implementation. Ask questions, clarify mid-build, or start planning what's next without interrupting anything.
 
-Named documents that carry context between agents with scoped visibility. An explorer writes findings, a planner reads them and writes a plan, a builder reads the plan and implements. Each agent sees exactly what it needs without inheriting everything.
+## Features
 
-### Two-way agent communication
+### Main agent → subagent architecture
+The main agent manages the conversation and delegates to specialized subagents (explorer, planner, builder, reviewer, browser, debugger), each with its own context window, toolset, and permissions. Subagents do focused work and report back, keeping the main agent's context clean.
 
-The orchestrator and subagents have full bidirectional messaging. The orchestrator can steer, redirect, or interrupt subagents mid-work. Not fire-and-forget delegation.
+### Two-way agent communication
+The main agent and subagents have full bidirectional messaging. The main agent can steer, redirect, or interrupt subagents mid-work. Subagents can message back when they hit blockers or need clarification. Not fire-and-forget delegation.
 
-### Event-sourced runtime
+### Context sharing via artifacts
+Shared markdown documents pass context between agents directly. An explorer writes findings, a planner reads them and produces a plan, a builder reads the plan and implements. No context is lost to summarization, and the main agent doesn't burn output tokens on handoff.
 
-Every action is an immutable event. Sessions are fully replayable and resumable.
+### Parallel by default
+The main agent spins up multiple subagents concurrently and keeps working while they run. Exploring separate areas of the codebase, implementing unrelated features, or debugging multiple issues all happen in parallel.
 
-### Built-in browser agent
+### Progressive disclosure of tool output
+Tool output does not enter the context window by default. Agents inspect results explicitly and pull in only what they need, keeping noisy output out of the conversation.
 
-A vision-based browser agent built on [browser-agent](https://github.com/magnitudedev/browser-agent) runs natively in the same runtime. Used to verify UI changes and behavior.
+### Sensible default permissions
+Commands are classified by risk. Read-only commands run freely. Normal commands are allowed in your project directory but blocked outside it. Dangerous commands are always blocked. No `--dangerously-skip-permissions`, no stuck permission prompts.
 
-### Steerable and hackable
+### Built-in browser agent
+A vision-based browser agent built on [browser-agent](https://github.com/magnitudedev/browser-agent) runs natively in the same runtime for verifying UI changes and behavior as part of the workflow.
 
-Skills, AGENTS.md, persistent memory, agent policies. The system adapts to how you work, not the other way around. Magnitude learns your preferences and codebase conventions over time, storing them in `.magnitude/memory.md` and applying them automatically to future sessions.
+### Persistent memory
+Magnitude learns your preferences and codebase conventions over time, storing them in `.magnitude/memory.md` and applying them automatically to future sessions. Skills and AGENTS.md are also supported.
 
 ## Additional Info
 
-### Provider and model support
+### Philosophy
 
-See the [providers](https://docs.magnitude.dev/configuration/providers) and [models](https://docs.magnitude.dev/configuration/models) pages in the docs. If you would like another provider to be supported, please feel free to create an issue or raise a PR yourself.
+The coding agent primitive is not solved. Most failures come from two recurring patterns: **context degradation** over long sessions, and **local maximum traps** where the agent settles for the nearest plausible fix. Magnitude is built around those failure modes. Read more in our [manifesto](https://magnitude.dev/manifesto).
 
 ### Contributing
 
@@ -93,5 +97,3 @@ Full documentation is available at [docs.magnitude.dev](https://docs.magnitude.d
 Built on top of [BAML](https://boundaryml.com), [Effect](https://effect.website), and [OpenTUI](https://github.com/anomalyco/opentui).
 
 Inspired by other open-source coding agents, including [OpenCode](https://github.com/anomalyco/opencode) and [Codex](https://github.com/openai/codex).
-
-The future is open!

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@magnitudedev/cli",
-  "version": "0.0.1-alpha.3",
+  "version": "0.0.1-alpha.4",
   "description": "Magnitude AI coding agent",
   "license": "Apache-2.0",
   "repository": {