Codex plugins landed in early 2026, and the directory filled up fast enough that most "top 10" posts are now a mix of official plugins, MCP servers in disguise, and skill packs that predate the plugin spec.
That mix is fine, as long as the article says which is which. I checked each entry below against OpenAI's curated directory or the maintainer's own repository, and I cite the roundup where another publication recommends the same thing.
| Plugin | What it is · Maintainer | Adoption signal · Also recommended by |
|---|---|---|
| GitHub | App connector · GitHub | 33,993 stars (MCP repo) · OpenAI curated directory |
| Figma | App connector · Figma | 16,077 stars (community MCP) · OpenAI curated, Firecrawl |
| Linear | App connector · Linear | hosted, no repo · OpenAI curated, Firecrawl |
| Sentry | App connector · Sentry | 948 stars (MCP repo) · OpenAI curated, AY Automate |
| Context7 | MCP + skill · Upstash | 63,751 stars · Firecrawl |
| Firecrawl | MCP + skills · Firecrawl | 7,698 stars (MCP repo) · Firecrawl |
| Superpowers | Workflow framework · Jesse Vincent | 1,009,371 installs (Claude marketplace) · Firecrawl |
| Matt Pocock's skills | Skill pack · Matt Pocock | Firecrawl reports ~125k stars (Jun 2026) · Firecrawl, aihero.dev |
| GitNexus | Code graph + MCP · Abhigyan Patwari | Firecrawl reports ~41.9k stars (Jun 2026) · Firecrawl |
| RTK | Output compression proxy · rtk-ai | Firecrawl reports ~61.3k stars (Jun 2026) · Firecrawl |
Two notes on that table. The install flow is /plugins inside Codex CLI, and @plugin-name to invoke one explicitly. And the star numbers attributed to Firecrawl are from their June 2026 roundup, not a live count, because I would rather label a number than pretend it is current.
The four official ones worth installing first
These come from OpenAI's curated directory, which means the vendor maintains the integration and the auth flow is managed for you.
1. GitHub
Review diffs, read issues, check workflow runs without leaving the session. The connector covers the parts of GitHub you actually use during a change, and it is the one I install on every machine.
Where it earns its place: CI failures on a branch I did not touch. The agent reads the run log and tells me which dependency changed, which used to be a fifteen-minute investigation.
2. Figma
Reads frames, layer names, text styles, and spacing from a real file. The output quality depends almost entirely on how the file is structured; named layers and auto-layout give usable code, loose frames give a first draft.
Firecrawl's roundup lists it, and the failure mode is worth repeating: updating an existing component often triggers full regeneration rather than a diff.
3. Linear
Pulls a ticket, updates its state, links the PR. The real benefit is continuity: a week later, the agent can summarize what shipped and what is blocked from actual issue state rather than your memory.
Depends on ticket quality. A vague ticket produces a vague implementation, and no plugin fixes that.
4. Sentry
Stack traces, breadcrumbs, release info, and grouping. 948 stars is low for how much time it saves, and the reason is boring: it is a vendor repository, not a community project that chased stars.
Pair it with the GitHub connector and an incident becomes three calls: which error, which release, which commit.
The four that change how the agent works
5. Context7
Version-specific library docs, fetched at call time. The highest-starred entry here, and the one that most directly reduces wrong API calls.
Firecrawl's roundup lists it, with the caveat I agree with: leave it on for every prompt and it burns context on libraries the model already knows. Enable it per project.
6. Firecrawl
Search, scrape, crawl, and map, returning clean markdown instead of raw HTML. Large results go to files, which keeps the context window from filling with boilerplate.
The plugin includes skills that teach Codex when to reach for each tool, which is the part most MCP servers skip. Without that, the agent picks scrape for a question that needed search.
7. Superpowers
A workflow framework rather than a tool: brainstorm before code, write a failing test first, investigate root cause before fixing, hand implementation to subagents with a review after each task.
It started in the Claude ecosystem and now ships Codex support with tool mappings of its own. 1,009,371 installs on the Claude marketplace, and Firecrawl lists it for Codex. The cost is verbosity: some sessions produce more planning text than code.
8. Matt Pocock's skills
A skill collection built around four failure modes: misalignment, verbosity, broken code, and architecture that turns into mud. The headliners are grill-me (interviews you before any code), tdd, diagnose, and handoff (compresses a session into a document for the next one).
It predates the plugin spec, installs with npx skills add, and shows up under "Created by you". Firecrawl reports about 125k stars, which would make it the most-starred skill pack in this list.
Two that fix context problems
9. GitNexus
Builds a knowledge graph of a repository in the browser and lets the agent query it with Graph RAG. The questions it answers are the ones file search handles badly: who calls this function, which modules depend on that one, what is dead code.
Client-side, so private repos stay private. The limit is size: very large monorepos push browser memory.
10. RTK
A Rust proxy between Codex and your shell that compresses command output before it reaches the context. git diff, kubectl describe, and failed install logs are the biggest token sinks in a normal session, and RTK reports 60 to 90 percent savings on routine output.
It works at the shell layer, so it also helps Claude Code, Cursor, and anything else that runs commands. The trade-off: compressed output sometimes drops a line you needed, and you have to know when to disable it.
The one I added myself
Every Codex setup I build ends up needing one domain tool that no marketplace covers. Mine is travel, so I wrapped an MCP server as a plugin and loaded it through config.toml.
[mcp_servers.rollinggo-hotel]
url = "https://mcp.rollinggo.ai/mcp"
http_headers = { Authorization = "Bearer YOUR_API_KEY" }
That gives Codex three tools: searchHotels, getHotelDetail, getHotelSearchTags. The reason to put it in Codex rather than a separate script is that the agent can check a price while it is editing the page that displays prices.
Honest evidence, since I have asked it of every other entry: 86 stars, MIT licensed, vendor-maintained, and no third-party roundup citation yet. The vendor reports 2,000+ developers and agents connected, 2M+ properties, and 110,000 direct contracts. It earns its slot on capability, not popularity, and the repository has the config, tool schemas, and response examples if you want to test that claim yourself.
How to choose, given ten options
Three questions, in order.
Is it maintained by the service owner? GitHub, Figma, Linear, Sentry, Context7 (for its own index), and Firecrawl all are. That matters because the API changes are their problem, not yours.
Does it replace a tab or a decision? A plugin that saves a tab is nice. A plugin that changes what the agent knows is worth the context it costs.
What does it add to every turn? Plugin descriptions load on every turn whether or not they fire. Four plugins is my ceiling for a coding session; beyond that, tool selection gets worse before it gets faster.
What I would skip
Autonomous "fix everything" plugins, anything that edits tests to make them pass, and collections with 200 skills where 190 never fire. The directory has enough good options now that volume is no longer a virtue.
If you want the setup I run, it is four official connectors, Context7, and one domain server. That is the whole config, and it covers 90% of a normal week.
Top comments (0)