OpenExpertise vs Anthropic /workflows

OpenExpertise is a stable, open-source orchestration layer you can use today; Anthropic /workflows is an unreleased preview feature baking the same "code-as-law" idea directly into Claude Code.

What /workflows is

Anthropic's /workflows is an unreleased feature surfaced in Claude Code v2.1.147–v2.1.148 and in a community preview repository (ray-amjad/claude-code-workflow-creator). It is not officially documented. Activation requires setting the environment variable CLAUDE_CODE_WORKFLOWS=1 and typing the keyword ultrawork inside a Claude Code session.

When triggered, Claude Code acts as an "architect agent": it generates a context analysis file, then writes a JavaScript control script (~300 lines for a real-world PR review) on the fly. That JS script is the workflow — a deterministic Node.js program that calls agent sub-processes inside its own loops. The model only runs inside explicitly defined agent() nodes; the outer control flow is pure code. This is the same "Code as Law" insight that OpenExpertise is built on.

The /workflows TUI is notably polished for a preview feature: a htop-style interactive dashboard shows each agent's status, elapsed time, live token count, and active MCP tool call. Generated scripts default to a 3-day ephemeral lifetime and must be explicitly promoted to a permanent path by the user. Anthropic's framing: LLM-generated scripts are "temporary assets" — test them, then promote.

Where they overlap

• Code-as-law philosophy — deterministic outer control flow, LLM confined to discrete nodes.
• One-keyword authoring — /workflows uses ultrawork; OpenExpertise uses oe ultra or /ultraexpertise inside Claude Code.
• TUI observability — both show real-time per-node status, token consumption, and active tool calls.
• Multi-agent fan-out — both can spawn parallel agent sub-processes within a single run.
• Claude Code as the underlying substrate — /workflows lives inside Claude Code; OpenExpertise can call Claude Code as a cli-agent node.

Where OpenExpertise differs

Dimension	OpenExpertise	Anthropic /workflows
Format	Declarative YAML — human-readable, diff-able, version-controlled	JS script generated per-run; not inherently VCS-friendly
Schema validation	oe validate checks AJV schema before any code runs	No pre-run validation; errors surface at runtime
Persistent state	SQLite blackboard, resumable via oe resume <run-id>	Ephemeral by default; 3-day TTL unless manually promoted
Self-evolution	oe evolve reads event log + state diff, proposes YAML patches	Not present
Multi-CLI	Calls Claude Code, OpenAI Codex, and Gemini CLI as node kinds	Anthropic-only
Multiple LLM providers	Anthropic + OpenAI + any OpenAI-compatible endpoint	Anthropic only
Availability	Stable, open-source, npm-installable today	Unreleased preview; no public ETA
Authoring artifact	A YAML file you own and evolve over time	LLM-generated JS script; may or may not survive 3 days
MCP server	Exposes 6 OE tools so Claude Code can orchestrate OE	Runs inside Claude Code; not callable externally

When to pick /workflows over OpenExpertise

• Your entire workflow lives inside Claude Code and you want zero context-switch to an external tool.
• You want to stay on native Anthropic tooling and prefer generated JS over hand-authored YAML.
• The Anthropic team ships /workflows GA with durability, versioning, and multi-provider support — at that point it may cover most of OE's surface for Claude Code users.

When to pick OpenExpertise over /workflows

• You need it today, with stability guarantees.
• You need multi-provider support (OpenAI, Gemini, vLLM, Ollama).
• Your SOP must survive across months, be version-controlled in git, and be reviewed by non-engineers.
• You need the evolution loop: run → observe → propose upgrade → apply.
• You want to call the orchestrator from outside Claude Code (via MCP server, programmatic API, or CI/CD).

Specific positioning

Both projects share the same core conviction — code controls flow, LLMs execute within nodes — but they serve different integration points. /workflows is Anthropic's first-party answer for Claude Code power users; OpenExpertise is the open-source answer for teams that need a stable, provider-agnostic, self-improving SOP runtime they can own.