bkit-codex plugin
An OpenAI Codex CLI extension: 27 skills, 16 MCP tools and AGENTS.md rules for PDCA development with a mandatory design document
- 72 ★ on the repository
- Last commit
- Apache-2.0
Extended access
Why this level
- The curl and bash installer writes files into the project and changes .codex/config.toml and AGENTS.md
- The MCP server runs on your machine and the process makes the agent write documents and edit code
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/ww-w-ai/bkit-codex/main/install.sh | bashRun in the project root. On Windows use install.ps1.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The curl and bash installer writes files into the project and changes .codex/config.toml and AGENTS.md
- The MCP server runs on your machine and the process makes the agent write documents and edit code
This is third-party code. Review the repository files before installing.
What it does
bkit-codex runs development in Codex CLI through the PDCA cycle: Plan, Design, Do, Check, Act and Report. Context is assembled in three layers. `AGENTS.md` rules (global and project, about 5.8 KB of 32) are always loaded, 27 skills load on demand and reference files open only when the agent asks for them. State is kept by 16 MCP tools on JSON-RPC 2.0 over STDIO with no external dependencies. The cycle has hard rules: you cannot jump from the plan straight to code and a design document is mandatory. Before code is written an MCP tool checks that the design document exists, and after significant edits (more than 10 new or 20 modified lines) it suggests a gap analysis. Check and Act repeat until the match rate reaches 90 percent, up to five iterations. After context compaction a `bkit_get_status` call with the `recovery` mode restores state. Commands are invoked as `$pdca plan`, `$pdca design`, `$pdca do`, `$pdca analyze`, `$pdca iterate` and `$pdca report`. The project level is detected automatically: Starter for static sites, Dynamic for full-stack apps with BaaS by signs like `docker-compose.yml` and `api/`, Enterprise for microservices by `kubernetes/` and `terraform/` directories. Each level has its set of phases from a nine-phase pipeline: schema, conventions, mockup, API, design system, UI integration, SEO and security, review and deployment. Separate skills cover code review, log-based zero-script testing and a Codex configuration guide.
Who it is for. Developers on Codex CLI who want a disciplined process with plan, design and report documents in place of free-form request-by-request coding.
Good fit when
- A new feature must go through a plan and a design document before code is written
- After an edit you want a gap analysis between design and code with iterations until a 90 percent match
- A project should be split into phases: schema, conventions, mockup, API, review and deployment
- After context compaction the agent loses PDCA state and it must be restored
Not a fit when
- You need Claude Code or Gemini CLI: the author keeps separate `bkit-claude-code` and `bkit-gemini` repositories for them
- You need hard automation through hooks: in the Codex version the agent follows rules itself, about 70 percent automation
- You need a quick one-file fix without documents: the process requires a design document before code
Example request
$pdca plan A user profile page with avatar upload, then create the design document with $pdca design.Limitations
The extension targets OpenAI Codex CLI 0.100.0 or newer and needs Node.js 20 and Git. Unlike the Claude Code version where hooks automate the process fully, here the agent reads `AGENTS.md` rules and calls MCP tools voluntarily itself, so automation is about 70 percent. The design document requirement is enforced by an MCP tool but is skipped if the agent does not call it.
The `install.sh` installer puts files into the project, configures the MCP server in `.codex/config.toml` and copies `AGENTS.md`. The license is Apache 2.0 and the last commit is from September 27, 2026.
How to disable. Remove the skill links from `.agents/skills/` and the `.bkit-codex` directory, take the MCP server block out of `.codex/config.toml` and the copied `AGENTS.md`. PDCA documents in `docs/` stay in the project and are deleted by hand.
FAQ
How do I install bkit-codex?
In the project root run `curl -fsSL https://raw.githubusercontent.com/ww-w-ai/bkit-codex/main/install.sh | bash`. Windows has `install.ps1`, and for a manual install clone the repository into `.bkit-codex`, link the skills, copy `AGENTS.md` and configure MCP in `.codex/config.toml`.
Where do I start on a feature?
With `$pdca plan` and a feature description, then `$pdca design`, `$pdca do`, `$pdca analyze`, `$pdca iterate` and `$pdca report`. `$pdca status` shows the status and `$pdca next` the next step.
What are the Starter, Dynamic and Enterprise levels?
They are project levels detected automatically. Starter is for static sites, Dynamic for apps with BaaS by `docker-compose.yml`, `.mcp.json` or `api/`, and Enterprise for microservices by `kubernetes/` and `terraform/` directories.
What do I do after context compaction?
Call `bkit_get_status` with `mode: "recovery"`: the tool restores the PDCA state and the task chain of the feature.
Is there a Claude Code version?
Yes, the author keeps separate `bkit-claude-code` and `bkit-gemini` repositories. There the process is automated by hooks, while in bkit-codex the agent follows the rules itself.
Discussion
No questions or reviews yet.
Related
PluginsA CLI that gives six coding agents project context, over 80 skills, verification and resumable tasks across sessions
A kit for porting a Claude Code workflow to OpenAI Codex CLI for multi-tenant PHP/Bitrix24-Marketplace projects
A Codex plugin adds development discipline, 29 skills, 31 hooks and multi-model subagents in one package.
A Claude Code plugin links planning, implementation and task checks, retains progress and controls how independently agents work.
61 small skills for Codex across 13 areas: codebase orientation, tests, migrations, security, CI and pull request review
Codex MCP Tool (cexll)
Codex MCP Tool
An MCP server that sends tasks to Codex CLI from Claude Code and Cursor: code review via @ refs, a sandbox, patch edits, ideas
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| bkit-codex (this card) | no | 72 | |
| Bkit | no | 605 | the Claude Code version with hooks, while here the Codex agent follows the rules itself through MCP |
| Context Engineering Kit | no | 1,821 | a set of context management techniques, while this one runs the PDCA cycle with a mandatory design document and state in MCP |
README in short
The description introduces context engineering as a way to manage agent context systematically and shows three layers: rules, skills and state in MCP. It then covers the architecture, commands, the PDCA cycle with hard rules, project levels, 16 MCP tools and installation.