Claude Code Source Code Leak: What the Unshipped Features Tell Us About Anthropic's Roadmap
Anthropic accidentally leaked Claude Code's source code. The hidden features—AutoDream, Chyros, Ultra Plan, and Coordinator Mode—reveal what's coming next.
Anthropic’s Accidental Reveal
When developers started poking around the Claude Code npm package in mid-2025, they didn’t find what you’d expect from a polished CLI tool. The source code was largely unminified — meaning the variable names, feature flags, and internal strings were all readable. Within hours, the AI community had catalogued a set of features that Anthropic had never announced publicly.
The Claude Code source code leak wasn’t a hack. Nobody broke in. The code was just… there. And what it contained offered a clearer picture of Anthropic’s product roadmap than anything the company had officially communicated.
This article walks through the four most significant discoveries — AutoDream, Chyros, Ultra Plan, and Coordinator Mode — and explains what each one suggests about where Claude and multi-agent AI development is heading.
How the Leak Happened
Claude Code ships as a Node.js package. Like most npm packages, the JavaScript inside is technically readable by anyone who installs it. The difference here was that the code wasn’t obfuscated or minified in a way that would obscure intent. Feature flags, string literals, and conditional logic were visible in plain form.
Developers who examined the package found references to features gated behind unreleased conditions — the kind of strings and flags that developers leave in codebases when something is built but not yet turned on for users.
This is a common pattern in software. Code gets shipped before the marketing is ready, before pricing is finalized, or before a feature is stable enough for public rollout. The difference is that most companies either minify aggressively or use obfuscation. Anthropic, at least for this package version, did not.
The result was an unintentional product announcement — made not through a blog post or a launch event, but through npm.
AutoDream: Autonomous Agent Loops
The name “AutoDream” appears in the source code in the context of what looks like an extended autonomous task mode — something beyond Claude Code’s existing ability to run multi-step coding workflows.
What the Code Suggests
AutoDream references point to a mode in which Claude Code could operate for extended periods without human interruption — executing, testing, revising, and iterating on code without waiting for approval at each step.
Current Claude Code behavior is already fairly autonomous. You can give it a task, and it will write files, run shell commands, and fix its own errors. But there are checkpoints — moments where it pauses to confirm before doing something consequential.
AutoDream appears to be designed to reduce or remove those checkpoints in specific contexts, enabling longer uninterrupted runs. Think of it as Claude Code shifting from “supervised autonomous” to “trusted autonomous” for low-risk or well-scoped tasks.
Why This Matters
This fits a broader pattern in agentic AI development. The most useful agents are ones that can handle extended tasks without hand-holding. But that requires trust — both in the model’s capabilities and in the guardrails around what it can and can’t do.
AutoDream suggests Anthropic is building toward a tiered autonomy model: different levels of human oversight depending on task type, risk level, or user trust settings. That’s a meaningful product design decision, and it signals that Claude Code is being positioned as something closer to a fully autonomous developer agent than a smart autocomplete tool.
Chyros: A New Model or a New Mode?
“Chyros” is the most opaque of the leaked references. It appears in conditional logic that seems to distinguish between different underlying models or execution modes — but the exact nature of Chyros isn’t fully clear from the source alone.
The Leading Interpretation
Most analysis points to Chyros being a codename for either a new Claude model variant or a specialized execution mode tuned specifically for code-heavy workloads.
Anthropic has historically used codenames during development before giving models their public names (e.g., internal names for Claude 3 variants before Haiku/Sonnet/Opus were announced). Chyros may follow the same pattern — a model in development that doesn’t yet have a public-facing name.
Alternatively, Chyros could be a distinct “mode” rather than a separate model — a configuration of Claude with different system prompt defaults, tool access, or context handling specifically optimized for software development tasks.
What It Would Mean
If Chyros is a model: it suggests Anthropic is working on a Claude variant specifically for coding, which would put it in direct competition with models like DeepSeek Coder, Qwen Coder, and the coding-specific fine-tunes that have proliferated in 2024 and 2025.
If Chyros is a mode: it suggests Claude Code is getting a dedicated execution layer — not just Claude with a coding system prompt, but a purpose-built experience with its own optimization profile.
Either way, the existence of Chyros as a distinct reference in the codebase means Anthropic is doing work here that hasn’t been publicly announced. That’s noteworthy on its own.
Ultra Plan: Above Pro
The “Ultra Plan” reference is probably the most commercially significant leak in the package.
What Was Found
The source code contains conditional logic that checks for an “ultra” plan tier — distinct from the existing Free, Pro, and Team plans. Users on this tier appear to have access to higher usage limits, and potentially to features (like AutoDream or extended agent runs) that aren’t available at lower tiers.
Pricing details aren’t visible in the code, but the plan structure is referenced clearly enough that it’s not ambiguous: Anthropic is building a higher-tier subscription above the current $20/month Pro plan.
Reading Between the Lines
This makes sense given market context. OpenAI launched a $200/month “Pro” tier (which it calls “ChatGPT Pro”) for power users who want higher rate limits and access to the most capable models. Google has similar premium structures for Gemini Advanced.
Anthropic has largely avoided the “super-premium for super-users” tier until now. The Ultra Plan reference suggests that’s changing — and that it’s being built alongside the heavier autonomous features like AutoDream, which would naturally require more compute per session.
For enterprise developers and serious Claude Code users, this is relevant. If Anthropic gates extended autonomous mode or higher-capacity agent runs behind Ultra, it creates a clear upgrade path for power users. It also signals that Anthropic sees a market segment willing to pay significantly more for more capable, less constrained Claude access.
Coordinator Mode: The Multi-Agent Signal
Coordinator Mode is the feature that carries the most long-term significance, especially for anyone building with Claude in multi-agent architectures.
What Coordinator Mode Appears to Be
References to “coordinator” and related logic in the source suggest a mode where Claude Code can act as an orchestrating agent — spinning up, directing, and synthesizing results from other agents or Claude instances.
This is the hub-and-spoke model of multi-agent AI: one agent coordinates the overall task, breaks it into sub-tasks, delegates to specialized sub-agents, then assembles the outputs into a coherent result.
In the context of Claude Code, this could mean:
- A single Claude Code session that spins up parallel sub-agents for different parts of a large codebase
- A coordinator that manages agents handling different languages, services, or repositories simultaneously
- Orchestration across tool-using agents that each specialize in specific tasks (database queries, API calls, test execution, etc.)
Why This Is Significant
Most current AI coding tools operate in a single-thread model: one model, one context, one conversation. Coordinator Mode suggests Anthropic is building Claude Code to operate as a multi-agent system natively — not just as a single agent that uses tools, but as an orchestrator that manages other agents.
This aligns with Anthropic’s published research on multi-agent systems, which has explored how to make agents more reliable by breaking complex tasks into smaller, more verifiable pieces.
For developers, this is the difference between “AI pair programmer” and “AI development team.” Coordinator Mode would let Claude Code tackle projects at a scale that a single-context agent simply can’t handle — large refactors, cross-service migrations, full-stack feature development where different components need to be worked on simultaneously.
What the Roadmap Signals
Taken together, these four features point to a consistent strategic direction for Anthropic.
Moving Toward Longer, More Autonomous Workflows
AutoDream and Coordinator Mode both point to the same thing: Anthropic is building Claude for tasks that take minutes to hours to complete, not seconds. This requires different infrastructure, different UX design (how does a user monitor an agent that’s running for 30 minutes?), and different safety considerations.
Tiered Access for Tiered Capability
The Ultra Plan, alongside the more capability-intensive features, suggests Anthropic is moving toward a model where raw capability is available at higher price points — not just higher rate limits, but access to qualitatively different features.
Claude as Infrastructure, Not Just Interface
Coordinator Mode in particular points to Claude Code being used not just as a developer tool but as infrastructure — a system that other agents and workflows plug into, rather than a standalone tool that a single developer interacts with.
This framing — Claude as an orchestration layer — is consistent with Anthropic’s push into enterprise and with the broader industry trend toward agents that connect to other agents, not just to humans.
Where MindStudio Fits Into This Picture
The features revealed in the Claude Code leak — particularly Coordinator Mode and multi-agent orchestration — point to a future where building with AI means building systems of agents, not just individual ones.
That’s exactly the pattern MindStudio is already built for.
Building Multi-Agent Workflows Without Waiting
You don’t have to wait for Coordinator Mode to go live to start building multi-agent workflows. MindStudio’s visual builder lets you connect multiple AI agents — each running different models, different prompts, different tools — into coordinated pipelines right now.
Where Claude Code’s Coordinator Mode would manage agent coordination inside a coding context, MindStudio handles it across any business process: research, content generation, data processing, customer support, API integration. You can build an orchestrator agent that routes tasks to specialized sub-agents, then synthesizes results — without writing the orchestration layer from scratch.
MindStudio supports Claude alongside 200+ other models, so if you want to use Claude as your coordinator and lighter models for sub-tasks (to optimize cost), that’s configurable out of the box.
The Agent Skills Plugin for Claude Code Users Specifically
For developers who are already using Claude Code and want to extend what it can do, MindStudio’s Agent Skills Plugin offers something directly relevant. It’s an npm SDK that exposes 120+ typed capabilities — sending emails, searching the web, generating images, running workflows — as simple method calls that any agent can use.
Claude Code, once Coordinator Mode ships, could use those methods to hand off tasks to external workflows mid-session. Until then, you can integrate the SDK into your own agentic tools today.
You can try MindStudio free at mindstudio.ai.
FAQ
What was the Claude Code source code leak?
The Claude Code npm package was published with largely unminified JavaScript, which made internal feature flags, strings, and conditional logic readable by anyone who installed the package. Developers discovered references to several unannounced features, including AutoDream, Chyros, Ultra Plan, and Coordinator Mode. Anthropic didn’t publicly comment on the specifics, but the code was genuine.
What is AutoDream in Claude Code?
AutoDream appears to be an extended autonomous mode that allows Claude Code to run longer task sequences without human confirmation checkpoints. It suggests Anthropic is building a tiered autonomy model — where users can opt into less-interrupted, more autonomous execution for appropriate tasks.
What is Coordinator Mode in Claude Code?
Coordinator Mode appears to be a multi-agent orchestration feature in which Claude Code acts as a coordinating agent — managing other agents or Claude instances, delegating sub-tasks, and synthesizing results. It’s designed for complex, parallel workflows that exceed what a single-context agent can handle.
What is the Claude Ultra Plan?
The Ultra Plan is an unreleased subscription tier referenced in Claude Code’s source code. It appears to be priced above the current Pro plan and is associated with higher usage limits and access to more advanced features. Anthropic hasn’t officially announced it, but conditional logic in the code clearly references a distinct “ultra” tier.
What is Chyros in Claude Code?
Chyros is a codename found in the Claude Code source code. It’s not entirely clear whether it refers to a new Claude model variant, a specialized execution mode, or something else. Most analysis suggests it’s either a forthcoming model tuned for code tasks or a distinct configuration mode for Claude Code with different default behaviors.
How does this affect developers building with Claude today?
The leaked features confirm that Claude Code is being built toward longer autonomous runs and multi-agent coordination — both significant for developers building agentic systems. In the near term, nothing changes. But understanding what’s coming lets you design your architecture now with those capabilities in mind, rather than retrofitting later.
Key Takeaways
- The Claude Code source code leak was accidental — the npm package shipped with unminified JavaScript that revealed internal feature flags and strings.
- AutoDream points to an extended autonomous mode with fewer human checkpoints during long task runs.
- Chyros is likely a codename for either a new code-focused Claude variant or a specialized execution mode.
- Ultra Plan confirms Anthropic is building a higher-tier subscription, likely tied to more compute-intensive autonomous features.
- Coordinator Mode is the most significant discovery — it suggests Claude Code will function as a multi-agent orchestrator, managing parallel agent workflows for complex tasks.
- Together, these features signal that Anthropic is positioning Claude Code as infrastructure for agentic development, not just a single-user coding assistant.
If you’re building AI-powered workflows today and don’t want to wait for these features to ship, MindStudio’s visual builder lets you build multi-agent systems with Claude and 200+ other models right now — no orchestration code required.