Codeflow is the command tower for AI engineering, and it already fits into modern coding-agent workflows, but the truthful entrypoint is
not “install our official plugin.” The truthful story today is:
read-only MCP, repo-owned skills, API/client entrypoints,
and Workflow Cases + Proof & Replay as the common operating record.
Use this page when you want the shortest honest answer to: “How do I actually use Codeflow with
Codex, Claude Code, or OpenClaw right now?”
The safe product positioning is workflow-aligned and read-only MCP-compatible.
Codex now has official plugin surfaces, Claude Code currently exposes project-level MCP / hooks / subagent surfaces,
and OpenClaw already has plugin + skills hosts of its own. Keep plugin wording narrow:
for Codeflow, the default public story is still integrations + skills + MCP plus local bundle examples, not a published listing.
Command Tower, Workflow Cases, Proof & Replay, read-only MCP, repo-owned skills, and API/client entrypoints all fit Codex-style governed coding loops.
Do not call Codeflow a published Codex directory entry or an OpenAI-managed integration surface.
Claude Code
Yes
The same repo-owned surfaces work for Claude Code workflows: read-only inspection, case tracking, operator guidance, and package-level builder glue.
Do not imply an official Claude Code plugin, Anthropic-managed partnership, or hosted account integration.
OpenClaw
Bundle-compatible only when explicitly mapped
OpenClaw belongs in the adjacent coding-agent layer. Codeflow can still provide read-only MCP, skill-pack guidance, and proof/case review surfaces around it without pretending the same first-class binding as Codex / Claude Code.
Do not present OpenClaw support as an official plugin, native OpenClaw plugin, or first-class product binding unless a real mapped bundle or native plugin ships and is tested.
Codeflow should wrap Codex workflows with Command Tower, Workflow Cases,
Proof & Replay, read-only MCP, and repo-owned skills or local bundle examples.
It should not be framed as a published Codex directory entry or an OpenAI-managed surface.
Codeflow should fit as the governed review/proof/control-plane layer around
those workflows. It should not be framed as a Claude Code marketplace listing
or Anthropic-managed plugin surface.
OpenClaw already has its own skills and plugin/catalog reality. Codeflow should
therefore land on the integration, proof, replay, and read-only MCP side first,
not pretend it already ships a native OpenClaw plugin.
Current distribution status by platform
Platform
Official surface exists?
Current Codeflow state
Do not overclaim
Codex
Yes. Official plugins surface exists.
Public starter docs, one local marketplace example, and one portable local bundle exist. There is still no official Codex listing receipt tracked here today.
Do not say Codeflow is listed in an OpenAI-managed directory unless you have the listing URL or receipt.
Claude Code
Yes. Official marketplace, submit, and review flow exist.
Public starter docs and a local --plugin-dir bundle seed exist. The OpenHands submission is filed separately, but there is still no Claude marketplace submission receipt tracked here today.
Do not say submitted, in review, or live without the matching official receipt.
OpenClaw
Yes. Official ClawHub registry exists.
Public starter docs plus a compatible local bundle path exist. The matching adoption-router skill is live on ClawHub.
Do not say in review for OpenClaw. Its official model is publish-first, not review-queue-first.
Do not imply a plugin marketplace or hosted account integration.
OpenClaw
OpenClaw repo + skills docs + ClawHub
Use Codeflow as the adjacent review/proof/read-only MCP layer, then map packages or skills only when a concrete workflow really needs them
Do not present Codeflow as a native OpenClaw plugin unless a mapped path is shipped and tested.
Three copy-paste starting points
If your team wants something more concrete than “read the docs,” start with one of these repo-owned paths.
# 1) protocol-first: prove the read-only MCP surface is real
bash scripts/run_codeflow_readonly_mcp.sh
# 2) builder-first: prove the thin client + contract bootstrap path is real
node packages/frontend-api-client/examples/control_plane_starter.local.mjs \
--base-url http://127.0.0.1:10000 \
--role WORKER \
--mutation-role TECH_LEAD \
--preview-provider cliproxyapi \
--preview-model gpt-5.4
# 3) proof-first: prove the public first-run loop is real
CODEFLOW_HOST_COMPAT=1 bash scripts/test_quick.sh --no-related
Need the client-specific setup files after that? Open the
agent starter kits for tracked Codex
config.toml, Claude Code .mcp.json, and
OpenClaw openclaw mcp set starter objects.
Portable drop-in recipe for coding-agent teams
If your real question is “What do I actually hand to Codex, Claude Code, or OpenClaw maintainers this afternoon?”,
keep the recipe small and honest:
Vendor the repo truth you want the agent to follow: README.md, AGENTS.md, and only the public route pages or specific skill folders you actually plan to reuse.
Wire the repo-owned MCP server as a stdio command, not as a hosted claim: bash /absolute/path/to/Codeflow/scripts/run_codeflow_readonly_mcp.sh.
Choose only one adoption lane first: protocol-first, skills-first, or builder-first. Do not promise plugin/store support unless a real platform surface exists and is verified.
# 1) keep the repo truth close to the agent
sed -n '1,220p' README.md
sed -n '1,220p' AGENTS.md
# 2) keep MCP on the truthful stdio path
bash /absolute/path/to/Codeflow/scripts/run_codeflow_readonly_mcp.sh
# 3) keep builder adoption copy-pasteable
node packages/frontend-api-client/examples/control_plane_starter.local.mjs \
--base-url http://127.0.0.1:10000 \
--role WORKER \
--mutation-role TECH_LEAD \
--preview-provider cliproxyapi \
--preview-model gpt-5.4
OpenClaw teams should use the same three lanes as an adaptation recipe, not as proof of a first-class native plugin.
How to talk about this truthfully
Good wording
“Codeflow works with Codex and Claude Code workflows through Workflow Cases, Proof & Replay, read-only MCP, repo-owned skills, and builder entrypoints.”
Bad wording
“Codeflow already ships the official Codex / Claude Code / OpenClaw listing.”
Good wording
“OpenClaw stays in the adjacent coding-agent layer and can use Codeflow as a bundle-compatible review surface when the mapped path is explicitly shipped and tested.”
Bad wording
“OpenClaw is a first-class officially supported plugin surface inside Codeflow today.”
FAQ
Is this an official Codex or Claude Code plugin page?
No. Codeflow is the command tower for AI engineering, not an official Codex or Claude marketplace listing. Codex now has official plugin surfaces and OpenClaw has official plugin and skills surfaces, but the current public story still ships workflow compatibility, read-only MCP, repo-owned skills, builder entrypoints, and local bundle examples rather than a published listing.
Where does OpenClaw fit today?
OpenClaw stays in the adjacent coding-agent layer. Codeflow can still provide read-only MCP, proof and replay, and skill-pack guidance around it without claiming a first-class native plugin.
Should a new team start from MCP, builders, or skills?
Start from read-only MCP when the first need is protocol inspection, from builders or API when the first need is package-level integration, and from skills when the first need is repeatable operating playbooks.
Does this page mean Codeflow is hosted or write-capable through MCP?
No. The current public MCP contract is read-only, and Codeflow should not be described as a hosted operator product today.