Skip to main content
plugin CX Codex

bindry — Claude Code plugin: what it does, how to install

Compiles Bindry Binders into task-loaded skills, either as pinned snapshots or through live MCP lookups. Useful when instructions should stay out of every prompt.

MIT ★ 1,596 updated 2026-10-29

Install into Claude Code · Plugin marketplace All methods and platforms ↓
/plugin marketplace add hashgraph-online/awesome-codex-plugins
/plugin install bindry@awesome-codex-plugins
Open on GitHub

What bindry does

bindry turns a public or workspace Bindry Binder into separate skills, one for each Skill it contains. pinned builds a local snapshot under .agents/skills/; live creates a pointer skill that asks the Bindry MCP tool bindry.skills.get for current content instead of using cached text.

The plugin provides $bindry-sync for compilation, $bindry-check for a read-only drift report, and $bindry-connect for one-time MCP setup on a machine. Codex uses local Node.js scripts and .agents/skills/; its plugin model exposes skills rather than a separate command layer.

Live mode needs real GUID Skill IDs, a Bindry workspace API key, and an MCP connection. A local Node.js executor is required: web-only ChatGPT chats do not run these scripts, so installing the plugin there does not activate them.

What it helps with

  • compile a Binder into .agents/skills/
  • load instructions only when a task matches
  • report drift in already compiled skills
  • fetch current Skill content through MCP

Example prompts for the agent

  • Run $bindry-sync to compile a Binder into .agents/skills/ with pinned mode.
  • Use $bindry-check to produce a read-only drift report for the compiled skills.
  • Run $bindry-connect to configure MCP for live mode.

How to install bindry

Plugin marketplace

/plugin marketplace add hashgraph-online/awesome-codex-plugins
/plugin install bindry@awesome-codex-plugins

Step by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own

Questions about bindry

How do I install bindry in Claude Code?

Run /plugin marketplace add hashgraph-online/awesome-codex-plugins, then /plugin install bindry@awesome-codex-plugins.

What does bindry do?

It compiles Bindry Binders into skills for Claude Code and Codex. The plugin also checks instruction drift and supports current content through MCP live mode.

What does bindry live mode require?

It requires real GUID Skill IDs, a workspace API key from Bindry, and an MCP connection using codex mcp add bindry --url <api-base>/api/mcp --bearer-token-env-var BINDRY_API_TOKEN. Attachments are still downloaded locally during compilation.

Source text

Plugins · MIT
# Bindry plugin for Codex CLI

Compiles a bound Bindry Binder into Codex skills — one skill per Skill — so instructions load only when a
task actually matches them, instead of sitting in context on every turn. Same idea as the Claude Code plugin
in this repo, ported to Codex's actual plugin/skill mechanics.

## Install

This plugin runs local Node.js scripts. It needs a local executor and is intended for Codex or local
threads in the ChatGPT desktop app; installing it does not make those scripts run in web-only chats.

```bash
codex plugin marketplace add bindryai/Plugins
codex plugin add bindry@bindry-plugins
```

Verified working end-to-end against the real `@openai/codex` CLI: both commands complete successfully and
`codex plugin list` shows `bindry@bindry-plugins` as `installed, enabled`, with all three skills and both
scripts present under the installed plugin's cache path.

## Directory package

The plugin manifest includes directory presentation metadata and the desktop-only declaration. From the
repository root, run `node scripts/package-codex.mjs <new-absolute-output-directory>` to build a standalone
bundle with its scripts, skills, icon, security policy and license at the package root. Stage newly added
source files first: the packager copies only Git-tracked files and excludes marketplace configuration.
ZIP the contents of that output directory for the OpenAI submission portal. Publisher verification and
directory review remain separate from building the bundle; a ZIP alone is not a published listing.

The marketplace manifest lives at `codex/.agents/plugins/marketplace.json` — this is the actual, confirmed
location (found by testing several guesses against the real CLI; **not** `.codex-plugin/marketplace.json`,
which is what a naive port from Claude Code's layout would suggest). The plugin manifest itself,
`.codex-plugin/plugin.json`, points its `skills` field at `./skills/`.

To try a local checkout instead of the published repo, point `marketplace add` at the repo root on disk
(`/path/to/Bindry.Plugins`) — same two commands otherwise.

## Use

Bindry ships three skills, each a near-direct port of the Claude Code plugin's equivalent command — same logic,
adapted to Codex's skill-only invocation model (Codex plugins have no separate "commands" concept; skills,
triggered by description-matching or `$name`, are the only invokable action a plugin can expose):

- **`bindry-sync`** — compile a Bindry Binder into `.agents/skills/` (Codex's project-local skills directory,
  the equivalent of Claude Code's `.claude/skills/`). Supports the same `--mode pinned|live` split described
  below.
- **`bindry-check`** — read-only drift report against the already-compiled skills. Never edits or re-syncs.
- **`bindry-connect`** — one-time, per-machine setup connecting Codex to the Bindry MCP server for live mode.

## Live mode

Same design as the Claude Code plugin: `--mode pinned` (default) compiles a static snapshot; `--mode live`
compiles a pointer skill whose body tells the agent to call the Bindry MCP server's `bindry.skills.get` tool
for current content instead of trusting anything cached. A live skill never goes stale, so `bindry-check`
reports it as "live (always current)."

Connecting to the MCP server is a deliberate, explicit opt-in via the `bindry-connect` skill — not a bundled
`mcpServers` entry in the plugin manifest. Codex, like Claude Code, has no way to conditionally load a
plugin-declared MCP server; bundling one would put a permanent "not connected" status in front of every
pinned-only installer who hasn't set up an API key.

```bash
codex mcp add bindry --url <api-base>/api/mcp --bearer-token-env-var BINDRY_API_TOKEN
```

Any key generated from the workspace's team page in Bindry works here — there's no separate permission to
grant for MCP access, and no separate token needed from pinned-mode sync. Verified live end-to-end: registered
a real API key this way against a real running Bindry API, confirmed `codex mcp get bindry` reports it enabled,
and confirmed (via a real independent MCP client) that `Authorization: Bearer <key>` against `/api/mcp` returns
real tool results — this required a small fix upstream in IBeam's default JWT scheme, which was previously
intercepting any non-JWT Bearer value before the API-key scheme got a chance; see the Bindry-API changelog for
that fix's own verification.

## Try it right now, without Codex

The compiler is the same plain, dependency-free Node script as the Claude Code plugin, just defaulting to a
different output directory:

```bash
# from a local export file
node scripts/compile-binder.mjs examples/git-flow-command-center.binder.json

# a Skill with an attachment (logo file)
node scripts/compile-binder.mjs examples/brand-guidelines.binder.json

# live, from a running Bindry API
node scripts/compile-binder.mjs <binder-id> --api-base http://localhost:5160 --token <api-key> --out .agents/skills
```

Verified against a real running Bindry API: output for the bundled examples is byte-identical to what the
Claude Code compiler produces for the same input (frontmatter, pin/live comment, body, assets — only the output
directory differs), and a full live round-trip (seed a Binder + Skill, compile pinned and live, edit the
Skill's content, re-fetch live and see the new content immediately) was verified the same way it was for the
Claude Code plugin.

## Current limitations

- `--mode live` requires real GUID Skill ids from a live-fetched Binder — the bundled
  `examples/*.binder.json` files use illustrative ids (e.g. `bnd_git_branch_pr_hygiene`), not real GUIDs, so
  those Skills are skipped with a warning rather than compiling a skill that would always fail.
- Attachments are still downloaded and pinned to disk at compile time even in live mode — the live MCP tool
  returns a Skill's text content only, not its file attachments.
- `/api/binders/{id}/export` is workspace-scoped, same as the Claude Code plugin — "sync your own Binders," not
  "install someone else's published Binder."
- Codex skills have no structured "prefer tool X" mechanism beyond the prose in the skill body itself (confirmed
  against a real OpenAI-authored plugin already installed on this machine) — a live skill's instruction to call
  the MCP tool is exactly that: an instruction, not an enforced constraint.

## Compiled skills need to be readable, and a failure here is nearly silent

Codex reads each `SKILL.md` **twice**, and both reads have to succeed for a skill to do anything:

1. **At session start**, Codex itself reads every `.agents/skills/**/SKILL.md` to build the list of skills the
   model is offered — name, description, path.
2. **During a turn**, the model reads the body of the skill it chose, with an ordinary shell command
   (`Get-Content` / `cat`) against that path. There is no skill tool in Codex's event stream; the body arrives
   this way or not at all.

So a compiled skill is unavailable if **either** Codex cannot read the file at startup, **or** the agent's shell
is sandboxed such that it cannot read the workspace during the turn. `codex exec -s read-only` is sufficient —
read-only permits reads — but a stricter sandbox would not be.

**The failure is almost invisible.** Observed directly (BIND-0264): with one `SKILL.md` made unreadable, Codex
emitted a single line to **stderr** —

```
ERROR codex_core::session::session: failed to load skill …\bulletin-protocol\SKILL.md: failed to read file: Access is denied. (os error 5)
```

— and then ran the turn normally. The skill was simply absent: it was never offered, never mentioned, and the
answer was written from general knowledge. Nothing in the response, and nothing on stdout, indicated that a
skill had been expected and lost. If you redirect or ignore stderr, there is no signal at all.

**The always-on block is unaffected.** In the same run the `AGENTS.md` preamble was still honoured in full, so
a Binder can be half-working — persona applied, skills silently missing — which looks like the Binder working.

If skills seem to be doing nothing on Codex, check stderr for that line before anything else. A malformed
`SKILL.md` fails the same way, for the same reason, with `missing YAML frontmatter delimited by ---`.

© Bindry · MIT · hashgraph-online/awesome-codex-plugins

Similar plugins

All plugins for Agent setup: memory, context, orchestration →
ub-mode unclebob The ub-mode skill for Claude Code helps developers structure code changes. It guides the development process by following the principles of "First Make it Work, Then Make it Right, Then Follow Practice." CX Codex +1 930installs methodology-advisor FlorianBruniaux The methodology-advisor command for Claude Code and Codex reads the codebase and asks three questions to recommend an AI-development methodology stack. Run it at the start of a project, before choosing a workflow. CX Codex +1 ★ 6,529repo stars discernment-nudge official anthropics discernment-nudge — a plugin for Claude Code and Claude.ai that appends two or three short questions after a substantive answer. Helps check key facts, probe the reasoning, and spot missing context. ★ 198,522repo stars example-skills official anthropics example-skills — a plugin for Claude Code and Claude.ai with a set of sample skills. Demonstrates how to build skills, MCP servers, artifacts, web tests, and internal communication tools. ★ 198,522repo stars impeccable pbakaus impeccable — a Claude Code plugin that gives the agent a designer's vocabulary and 24 commands for working with interfaces. Kick it off at any stage, from gathering product context to the final polish pass. ★ 114,575repo stars superpowers obra superpowers — a plugin for Claude Code and Codex that teaches the agent a complete methodology: requirements gathering, design, TDD, and subagent-driven development. Triggers automatically when you start building something new. CX Codex ★ 323,788repo stars ponytail DietrichGebert ponytail is a Claude Code plugin that steers coding tasks toward short, practical solutions. Its review command checks related code for bugs, security issues, and missing tests. ★ 196,420repo stars last-30-days-skill mvanhorn last-30-days-skill searches the web for what people posted about a topic in the last month and turns it into one structured brief. ★ 72,053repo stars

From the same set391

The whole set →
chatgpt-agent chatgpt-agent is a Claude Code and Codex plugin that brings ChatGPT into review, planning, editing, and implementation tasks. It uses ZIP, MCP, or GitHub access. CX Codex ★ 1,596repo stars casefile casefile tracks persistent security findings for bug bounties and CTFs with isolated Docker PoC validation. CX Codex ★ 1,596repo stars bringyour-migration-auditor bringyour-migration-auditor runs a read-only check on a project that just moved from Claude to Claude: rule files, hooks, MCP config, skills, secret references. Use it before the agent edits code. CX Codex ★ 1,596repo stars boss boss chains nine specialist roles from requirements to deployment in Claude Code and Codex. Enforces strict quality contracts and writes disk artifacts. CX Codex ★ 1,596repo stars bgs-modding-superpowers bgs-modding-superpowers — a plugin for Bethesda modpack curation: two MCP servers, an xEdit conflict audit, and agent control over your MO2 install. CX Codex ★ 1,596repo stars babok_analyst babok_analyst guides business analysis across 9 BABOK v3 stages with mandatory human sign-off at every gate. CX Codex ★ 1,596repo stars aw1-circuit-breaker The aw1-circuit-breaker plugin checks ASTs and tool-call arguments before Python runs code. Use it to intercept dynamic execution, shell launches, and socket connections. CX Codex ★ 1,596repo stars auto-pilot auto-pilot turns a short Claude request into a one-question-at-a-time intake, saves the answers as project files, and lets a later session pick up where the last one stopped. CX Codex ★ 1,596repo stars

Project sponsors

Companies that keep this analytics open