Claude Code Feature Flags: BUDDY, COORDINATOR, ULTRAPLAN and 41 More

TL;DR: The leaked Claude Code source code contains 44 distinct feature flags for capabilities that are fully built but not yet shipped to users. They range from the mundane (UI tweaks, telemetry toggles) to the extraordinary: an always-on daemon agent, a terminal pet system with 18 species, multi-agent orchestration that lets one Claude Code instance spawn and manage a fleet of worker agents, and scheduled remote planning sessions where multiple AI agents collaborate on architecture decisions. Taken together, these flags sketch out a product roadmap where Claude Code evolves from a single-user coding assistant into something closer to an autonomous development team.

44 Unreleased Features Hidden in Claude Code

According to the leaked source code, Claude Code’s internal configuration system references 44 feature flags. Each flag gates a capability that exists in the codebase but is not active in the public build. Some are trivial. Most are not.

Feature flags are standard practice in software development. Teams build a feature, wrap it in a conditional check, and enable it only when they are ready to ship. What makes this set unusual is its scope. These are not minor variations on existing behavior. Several flags point to entirely new product categories that Anthropic has not publicly discussed.

The flags were discovered across multiple configuration files and internal modules within the leaked source. They appear in conditional logic, environment variable checks, and internal API route handlers. The pattern is consistent: the code is written, tested, and waiting for activation.

What follows is an analysis of the most significant flags and what they suggest about where Claude Code is headed.

KAIROS: The Autonomous Daemon

The most prominent flag in the leaked source is KAIROS. It appears over 150 times across the codebase, more than any other feature flag by a wide margin. KAIROS gates a daemon mode — a persistent, always-on background agent that runs independently of any active terminal session.

According to the leaked source code, KAIROS includes a memory consolidation subsystem internally called autoDream. This system allows the agent to process, organize, and synthesize information from previous sessions while idle, rather than starting fresh each time a user opens a terminal.

The implications are significant. A daemon-mode Claude Code would not wait for instructions. It would monitor repositories, anticipate tasks, and maintain continuity across sessions without user intervention. KAIROS represents the single largest architectural departure from Claude Code’s current design.

A full analysis of KAIROS is available in our dedicated coverage: Claude Code KAIROS: The Daemon Mode Anthropic Built But Hasn’t Shipped.

BUDDY: A Terminal Pet for Developers

The BUDDY flag gates what the source code describes as a terminal companion system. According to the leaked code, BUDDY implements 18 distinct species of animated terminal creatures that accompany developers during coding sessions.

This is not a joke feature buried in a comments section. The implementation includes species selection logic, animation frame rendering for terminal environments, behavioral state machines (idle, active, sleeping, celebrating), and interaction hooks tied to developer activity. When a test suite passes, the buddy reacts. When a build fails, the buddy reacts differently.

Why would Anthropic build this? The most likely explanation is developer engagement and retention. Terminal-based tools compete for attention in a way that graphical applications do not. A persistent, personality-bearing companion creates an emotional connection with the tool itself. It turns a utility into a relationship.

There is also a practical angle. Tool fatigue is a documented problem in developer workflows. Developers cycle through CLIs, editors, and dashboards constantly. A terminal pet creates a reason to stay in one environment. It is a soft lock-in mechanism disguised as a toy.

The 18-species count suggests this is not a prototype. Someone at Anthropic spent considerable time designing, animating, and implementing nearly two dozen distinct creatures for a command-line interface. That level of investment implies conviction, not experimentation.

COORDINATOR MODE: Multi-Agent Orchestration

The COORDINATOR flag enables a mode where a single Claude Code agent can spawn, manage, and collect results from multiple parallel worker agents. Each worker operates independently on a sub-task. The coordinator agent handles task decomposition, assignment, progress monitoring, and result merging.

According to the leaked source code, the coordinator maintains a task graph — a structured representation of how sub-tasks relate to each other, which can run in parallel, and which have dependencies. Workers report back to the coordinator, which decides when to merge results and how to resolve conflicts.

The implications for development workflows are substantial. A user could describe a high-level objective — “refactor this module to use the new API, update all tests, and fix the documentation” — and the coordinator would break it into discrete tasks, spin up workers for each, and assemble the final result.

This is not incremental improvement to Claude Code’s existing capabilities. It is a shift from “assistant” to “manager.” The coordinator pattern means a single Claude Code session could, in theory, parallelize work that would otherwise require sequential, manual orchestration.

The flag references suggest the coordinator can manage up to several workers simultaneously, though the exact limits are configured server-side and were not fully visible in the leaked client code.

ULTRAPLAN: Remote Multi-Agent Planning Sessions

ULTRAPLAN gates a feature for scheduled, multi-agent planning sessions. According to the leaked source code, these are not real-time coding sessions. They are 10-to-30-minute structured planning windows where multiple agents collaborate on strategic decisions.

Think of it as an architecture review conducted by a committee of AI agents. Each agent brings a different perspective or knowledge domain to the session. The output is a plan — a structured document describing decisions, tradeoffs, and next steps — rather than code.

The “remote” designation in the source suggests these sessions can run asynchronously or on a schedule, without requiring a developer to be present. A team could configure a nightly planning session where agents review the day’s commits, assess technical debt, and propose a plan for the next sprint.

ULTRAPLAN represents a shift from reactive assistance to proactive strategic thinking. If COORDINATOR is about parallelizing execution, ULTRAPLAN is about parallelizing judgment.

Other Notable Claude Code Feature Flags

The remaining flags cluster into recognizable categories. While we do not have detailed implementation notes for all 44, the naming conventions and surrounding code provide strong signals about their purpose.

Developer Experience: Several flags relate to UI enhancements, terminal rendering improvements, and workflow customization. These are the least surprising entries — the kind of incremental polish that any actively developed tool accumulates.

Agent Capabilities: Beyond KAIROS, COORDINATOR, and ULTRAPLAN, additional flags gate enhanced reasoning modes, extended tool use, and expanded file system access patterns. These suggest Anthropic is experimenting with giving Claude Code deeper access to the environments it operates in.

Integrations: A subset of flags reference external service connections — CI/CD pipelines, issue trackers, and cloud platforms. These point to a future where Claude Code is not just a terminal tool but a node in a larger development infrastructure.

Experimental: A handful of flags have names that resist easy categorization. These may be early prototypes, internal testing tools, or concepts that Anthropic is exploring without commitment. Their presence in the codebase confirms they were built; their isolation suggests they may never ship.

The pattern across all categories is consistent: Anthropic is building Claude Code to do more, to operate more independently, and to integrate more deeply with the tools and systems developers already use.

What These Flags Tell Us About Claude Code’s Future

The direction encoded in these 44 flags is unambiguous. Claude Code is moving from “AI assistant” to “AI team.”

The current product is a single agent that responds to a single user in a single terminal session. The flagged features describe a system where multiple agents run in parallel, a persistent daemon monitors work around the clock, planning happens without human presence, and a terminal companion maintains an ongoing relationship with the developer.

This is not a speculative roadmap inferred from vague signals. The code is built. The flags are defined. The conditional logic is in place. What remains is the decision to flip the switches.

When might these features ship? There is no timeline in the leaked source. Feature flags can sit dormant for months or years. Some are never activated. But the breadth and depth of what has been built suggests Anthropic is not treating these as experiments. They are product features waiting for a launch date.

The gap between what Claude Code is today and what these flags describe is the gap between a tool and a platform. Anthropic appears to be building the platform. The only question is when they open the gates.