Best Codex Plugins in 2026

The best Codex plugins in 2026, going by HOL's scan data, are avoid-ai-writing, Stop That Shit, Better Harness, deja, YouTube Skills by TranscriptAPI, Codex Autoresearch, MisakaNet, Claude Code for Codex, Lattice and Nightshift. Each one has a Safe security label in the HOL Plugins catalog, no high or critical findings, and an install command we checked against the plugin's own README.

The Best Codex plugins page on hol.org ranks the full Codex catalog by Trust Score. This post answers a narrower question: if a plugin passed the scan, which ones have the most developers behind them, and what does each one do for you inside Codex? We took the 618 Codex listings, kept those with a clean scan, and picked the most-starred plugin for each of ten common Codex jobs, so the list doesn't hand you three tools that do the same thing.

TL;DR: best Codex plugins

Trust Scores were read from the HOL catalog on Oct 27, 2026. Star counts come from HOL's last GitHub fetch on Oct 14, 2026, so the live numbers on GitHub may be higher.

How we picked these

HOL indexes 618 plugins that list Codex as a supported runtime. 149 of them carry a Safe security label, which means the security part of the score is 90 or higher. 33 have not been scanned yet, and 4 are labeled Unsafe.

From the Safe group we kept plugins with a Trust Score of 78 or more. Then we chose ten jobs people hand to Codex (prose cleanup, scope control, workflow review, session memory, video research, performance work, failure fixes, second-model review, design discipline and long runs) and took the most-starred plugin for each. A plugin that repeated a job already covered was skipped, which is why humanizer-ru, a Russian-language cousin of avoid-ai-writing, isn't here. Clean plugins for other jobs, such as B2B prospecting or cloud deployment, may appear in later roundups. For each survivor we opened the README and copied the Codex install steps exactly as the maintainer wrote them.

The Trust Score methodology is public. It is a weighted sum of security (40%), provenance (20%), installability (15%), MCP posture (10%), publisher (10%) and maintenance (5%), capped hard when the scan finds a high or critical issue. Paid placements and memberships are not inputs. The scan is static: it reads source files and manifests and does not run the plugin. A high score is evidence, not a promise, so read the permissions before you install.

What are Codex plugins?

A Codex plugin is a package that adds capabilities to Codex in one install. According to OpenAI's plugin docs, a plugin can contain any mix of four parts:

• Skills: reusable instructions Codex loads when a task matches, often with reference files or helper scripts.
• MCP servers: connections to outside tools and data, with their own authentication.
• Browser extensions the plugin's workflow needs.
• Hooks: commands that run at set points in the Codex runtime. Codex asks you to review and trust them before they run.

ChatGPT and Codex share one public plugin directory, so a plugin published there shows up in both. Plugins work in Codex inside the ChatGPT desktop app and in Codex CLI. The IDE extension does not support plugins.

How to install Codex plugins

In the ChatGPT desktop app

Open the Plugins tab, search for the plugin, open its details and select the plus button. If the plugin needs an MCP connection, connect it when prompted. Then start a new chat. Type @ to call a plugin or one of its skills by name, or describe the task and let Codex pick the tool.

In Codex CLI

Run /plugins inside a Codex session to open the plugin browser. It groups plugins by marketplace, and you press Space on an installed plugin to turn it on or off. Most community plugins live in a GitHub marketplace, which you register once from your terminal:

codex plugin marketplace add owner/repo
codex plugin add plugin-name@marketplace-name

Use --ref to pin a tag or branch. Start a new session after installing so the bundled skills and tools load.

Skill-only packages

A few entries below ship as Agent Skills rather than a full plugin. Codex reads skills from .agents/skills/ in a repo or ~/.agents/skills/ for every project, and many projects install there with npx skills add owner/repo.

Codex skills vs Codex plugins

In practice the line is thin. Several plugins on this list are mostly skills wrapped in a plugin manifest so they can be versioned and updated through a marketplace.

The 10 best Codex plugins

1. avoid-ai-writing: cut AI tells out of your prose

Conor Bronsdon's skill audits text for AI writing patterns and rewrites it. It has a rewrite mode, a detect-only mode, and an edit mode that changes a prose file in place. The optional local detector calls no model and sends no text to a network service.

• Trust Score 80, security 95 (Safe), 0 high findings
• 4,993 GitHub stars, 445 forks
• 9 skills, owner verified on HOL

Install: it is listed in the OpenAI Plugins Directory, so search for it in the Plugins tab. To use it as a plain skill instead:

git clone https://github.com/conorbronsdon/avoid-ai-writing .agents/skills/avoid-ai-writing

Example prompts from the README:

Remove AI-isms from this post
Audit this draft for AI tells

Pros: the most-starred clean-scanning Codex listing in the catalog. Detect mode lets you see what it would flag before it touches anything.

Cons: edit mode refuses source code, configuration and generated data, so it is for prose only. The full rule set is large; the single-file version is about 138 KB, which costs context if you paste it into a rules field.

HOL listing · GitHub

2. Stop That Shit: keep Codex inside the task you gave it

A local guard for the habits that waste tokens: refactoring modules you didn't mention, adding checksums nobody reads, editing files during a read-only review, and re-running checks that already passed. You set a mode on the first line of a prompt, and its hooks block covered actions that fall outside it.

• Trust Score 92, security 100 (Safe), 0 high findings
• 2,616 GitHub stars
• 2 skills plus Codex hooks

Install (pinned to the 0.2.5 tag, as the README recommends):

codex plugin marketplace add lennney/stop-that-shit --ref 0.2.5
codex plugin add stop-that-shit@stop-that-shit

Restart Codex, then type /hooks in a fresh session to review and trust the hook commands.

Example prompts from the README:

$stop-that-shit review -- Review this diff. Report findings; do not edit.
$stop-that-shit change -- Fix the config read failure, including affected callers and checks.

Pros: the highest Trust Score on this list. Per-task overrides such as deps=allow or hash=allow let you loosen one rule without turning the guard off.

Cons: the main README is in Chinese, with an English README linked from it. Hooks need a manual trust step after install. The listing is not owner verified on HOL yet.

HOL listing · GitHub

3. Better Harness: find out what slows your agent down in this repo

Qoder's plugin reads project and session signals and produces a report of prioritized, verifiable changes to how your repository works with coding agents. It covers session evidence, project setup, agent customization and loop design.

• Trust Score 81, security 100 (Safe), 0 high findings
• 2,377 GitHub stars, 192 forks
• 11 skills; also packaged for Claude Code, Cursor and Kimi

Install:

codex plugin marketplace add https://github.com/QoderAI/better-harness.git --ref main
codex plugin add better-harness@better-harness

Example prompt from the README:

$better-harness:better-harness analyze this project's AI coding workflow and generate an evidence-backed report

Pros: the output is a report with evidence attached, so you decide what to change. Useful on a repo where several people use different agents.

Cons: the catalog version is 0.7.0-alpha2, so expect changes. The README documents one Codex prompt; everything else happens inside the report flow.

HOL listing · GitHub

4. deja: give Codex the sessions you already had

Codex keeps its session logs in ~/.codex/sessions, and Claude Code, Cursor and dozens of other agents keep their own. deja indexes all of them locally, including sessions from before you installed it, and hands relevant past work to the agent at session start, before an edit, or after a command fails.

• Trust Score 87, security 95 (Safe), 0 high findings
• 1,190 GitHub stars
• 1 skill and 1 MCP server, owner verified on HOL

Install the binary and wire it into every agent it finds:

brew install deja-vu
deja install --auto

Or install only the Codex package:

codex plugin marketplace add https://github.com/vshulcz/deja-vu
codex plugin add deja-vu@deja-vu

Example usage from the README:

deja "connection pool exhausted"
deja handoff --to codex

Pros: no model and no server in the loop, and keys and tokens are stripped while the index is built. A fix found in Codex comes back in Claude Code and the other way around.

Cons: HOL lists it primarily as a Claude Code plugin; the Codex package is one of several. On Windows the shell installer does not work, so use Scoop or winget.

HOL listing · GitHub

5. YouTube Skills by TranscriptAPI: read videos without watching them

Twelve skills for YouTube transcripts, timestamped captions, video and channel search, playlists and channel browsing. Most people only need youtube-full, which covers all of it in one skill.

• Trust Score 87, security 95 (Safe), 0 high findings
• 1,059 GitHub stars, 120 forks
• Skill pack (HOL's scan indexed 24 skill files)

Install:

npx skills add ZeroPointRepo/youtube-skills --skill youtube-full

Example prompts from the README:

Summarize this video: [URL]
Search the TED channel for 'artificial intelligence'

Pros: plain-English prompts with no configuration. The free tier gives 100 credits with no card.

Cons: every call goes through the hosted TranscriptAPI service. First run asks for your email and a one-time code to create the API key, and heavy use will need paid credits.

HOL listing · GitHub

6. Codex Autoresearch: make code faster with measurements, not guesses

You give it a benchmark, correctness checks, an edit boundary and a time budget. It records a baseline, tries small changes, repeats measurements before keeping one, and leaves a reviewable patch with the evidence. Version 3.0 adds bounded investigations where you know the outcome but not the method.

• Trust Score 83, security 100 (Safe), 0 high findings
• 834 GitHub stars
• 1 skill, Codex only

Install from /plugins by choosing TheGreenCedar, then codex-autoresearch, then Install plugin. To register the marketplace from your terminal first:

codex plugin marketplace add TheGreenCedar/AgentPluginMarketplace --ref main

Example prompts from the README:

/goal @Codex Autoresearch run a measured loop to reduce parser runtime.
/goal @Codex Autoresearch help me design a trustworthy benchmark for my indexer's speed and memory use.

Pros: Codex proposes the full contract (workload, checks, scope, budget) and waits for your approval before it runs anything.

Cons: narrow by design; reviews, docs and one-off fixes stay ordinary Codex work. The maintainer states that no model-driven comparison has been run, so there is no claim it beats plain Codex on cost or results.

HOL listing · GitHub

7. MisakaNet: stop debugging the same failure twice

A git-backed library of failure lessons for coding agents. When an error appears, the agent searches it; if a verified fix exists, it applies it, and if not, the miss becomes an intake for a new lesson. Every lesson is a Markdown file, reviewed like code and graded from E0 (intake) to E4 (reuse observed by two or more distinct clients).

• Trust Score 83, security 100 (Safe), 0 high findings
• 525 GitHub stars, 217 forks
• 1 skill plus a public MCP endpoint, owner verified on HOL

Install (the setup tool writes MisakaNet into Codex's config):

npx @misaka-net/misakanet-setup
npx @misaka-net/misakanet-setup --verify

Example: the README's own first test is a search for the error text database is locked. In Codex, paste an error you have hit before and ask it to check MisakaNet for a known fix.

Pros: no account or token needed. The evidence grade tells you how much to trust a fix before you run it.

Cons: the corpus leans toward the domains its contributors work in (RAG, DevOps, Docker, CI, WSL, Windows and others), so coverage of your stack may be thin. The hosted endpoint adds a semantic fallback when keyword search misses; set MISAKANET_SEMANTIC=0 to turn it off.

HOL listing · GitHub

8. Claude Code for Codex: a second model's review without leaving Codex

Sendbird's plugin turns Codex into a host for Claude Code jobs. You get seven commands for reviews, adversarial reviews that question the design, rescue tasks that hand a bug to Claude Code, and tracked background jobs whose results come back into the Codex thread.

• Trust Score 83, security 100 (Safe), 0 high findings
• 222 GitHub stars
• 7 skills plus Codex hooks

Install, then run setup once:

codex plugin marketplace add sendbird/codex-marketplace
codex plugin add cc@sendbird

Example prompts from the README:

$cc:review --base main
$cc:rescue investigate why the tests started failing

Pros: a different model checks Codex's work, and the optional review gate at the end of a turn is off until you enable it with $cc:setup --enable-review-gate.

Cons: you need the Claude Code CLI installed and signed in, plus Node.js 18+ and Codex hook support, so you need access to both products.

HOL listing · GitHub

9. Lattice: make Codex design before it codes

Rahul Garg's framework splits engineering discipline into small skills: atoms for single principles (clean code, architecture, DDD, secure coding, test quality), molecules that chain them into design, implement, refactor, fix and review workflows, and refiners that interview you to set project standards. Decisions and review notes collect in a .lattice/ folder that the next feature reads.

• Trust Score 78, security 95 (Safe), 0 high findings
• 204 GitHub stars
• 60 skills, owner verified on HOL; also packaged for Claude Code, Cursor, Grok and Kimi

Install:

codex plugin marketplace add techygarg/lattice
codex plugin add lattice@lattice

Example commands from the README, typed in the Codex chat:

/lattice-init
/design-blueprint

Pros: the pipeline from requirement to design to code to review produces files a team can check into the repo, not one person's prompts.

Cons: Codex has no subagent concept, so Lattice's verification runs a script instead of a reviewer agent. Walking five design levels before any code is slow for small changes. It has the lowest Trust Score on this list.

HOL listing · GitHub

10. Nightshift: hand Codex a shift and review the receipts

Nightshift gives long Codex runs a work contract. You bring a task list or an objective and a number of hours, and a clock-out gate rejects early stops while items remain. Progress survives context compaction on disk, decisions go to a parking lot with a default, and you get per-item receipts with token usage and time where the host reports them.

• Trust Score 89, security 95 (Safe), 0 high findings
• 76 GitHub stars
• 15 skills, owner verified on HOL, MIT licensed

Install from the OpenAI Plugin Directory in the Plugins tab, or from its marketplace:

codex plugin marketplace add orwa-mahmoud/nightshift
codex plugin add nightshift@nightshift

Example prompts from the README:

Set up Nightshift in this project.
Start the Nightshift shift.

Pros: listed in OpenAI's own directory, needs no separate service or API key, and runs on macOS, Linux and native Windows.

Cons: the youngest project here by stars. Unattended multi-hour runs carry risk, and the README asks you to keep the first run attended and work through its first-night checklist. A disposable ChatGPT conversation can't keep project changes, so it needs a real repo or folder in Codex.

HOL listing · GitHub

Popular Codex plugins we left out

Stars alone would have produced a different list. Ponytail, a skill that pushes agents toward minimal solutions, has 162,795 stars but no Trust Score yet because HOL hasn't finished scanning it. Planning with Files has 27,433 stars, but its security score of 59 puts it in the Caution band. Agent402 has a Trust Score of 94, among the highest in the Codex catalog, with 41 stars; it would make a trust-first list, which is what the Best Codex plugins page is for.

Build your own Codex plugin

The fastest start is the built-in plugin creator: type $plugin-creator in Codex (or @plugin-creator in ChatGPT Work mode) and describe the workflow. It scaffolds a .codex-plugin/plugin.json manifest and can add a local marketplace entry so you can test the plugin right away. OpenAI's build guide also documents a portable layout with a root plugin.json, a skills/ folder and an mcp.json for bundled MCP servers.

To test from a folder on disk, point the CLI at it:

codex plugin marketplace add ./local-marketplace-root

Before you share it, run the HOL Plugin Scanner on the repo. It checks Codex packages for leaked secrets, unsafe MCP settings, risky approval defaults and broken manifests, and it runs in GitHub Actions. Then submit the plugin to HOL Plugins so it gets a public Trust Score and a listing other Codex users can find. If you want a check on what plugins do at runtime on your own machine, HOL Guard reviews agent actions before they run.

FAQ

What is the best Codex plugin in 2026?

It depends on the job. Among Codex plugins with a clean HOL scan, avoid-ai-writing has the most GitHub stars (4,993) and Stop That Shit has the highest Trust Score on this list (92). For a list ranked purely by Trust Score, see the Best Codex plugins page.

Are Codex plugins safe to install?

Not automatically. A plugin can bundle hooks and MCP servers that run commands or reach outside services. Check the plugin's HOL listing for its security label and findings, read its README and permissions, and review hooks when Codex asks you to trust them. HOL's scan is static and is not a safety guarantee.

How do I install a plugin in Codex CLI?

Run /plugins in a Codex session to browse configured marketplaces. For a plugin hosted on GitHub, run codex plugin marketplace add owner/repo, then codex plugin add name@marketplace, and start a new session.

Do Codex plugins work in the VS Code extension?

No. OpenAI's docs say the IDE extension does not support plugins. Use Codex in the ChatGPT desktop app or Codex CLI.

What is the difference between a Codex skill and a Codex plugin?

A skill is one folder of instructions with a SKILL.md file. A plugin is a package that can bundle several skills plus MCP servers, hooks and assets, and installs through a marketplace.

How is the HOL Trust Score calculated?

It is a 0 to 100 weighted score from a static scan: security 40%, provenance 20%, installability 15%, MCP posture 10%, publisher 10% and maintenance 5%. Any high or critical finding caps it at 39. The full method is on the Trust Score page.

How do I get my plugin on this list?

Submit it through HOL Plugins, fix anything the scan flags, and keep a README with clear Codex install steps. We refresh this list from the catalog data.