Sendbird Client Plugin vs. Other Chat SDK Skills for Claude Code and Codex

The Sendbird Client Plugin is Sendbird's official plugin for Claude Code and Codex. It helps your coding agent add Sendbird Chat, Chat UIKit, and 1:1 voice and video calls to iOS, Android, React, and React Native apps. It also ships hooks: scripts that check the agent's work before a file is written and before an answer is returned.
Several chat and real-time vendors now publish agent skills for AI coding tools. This post compares the Sendbird Client Plugin with the public agent-skill repositories of five of them: Stream, CometChat, Twilio, Agora, and TalkJS. It looks at what each one sets up, how each one guards against invented APIs, and how each one checks that the app works.
Read the Client Plugin guide, start with Add Chat, or browse the plugin source on GitHub.
When to consider the Sendbird Client Plugin
Consider it when you build Chat or 1:1 voice and video calls into an iOS, Android, React, or React Native app, and you want SDK guidance plus hooks that check supported changes before they are written. Its scripted run check helps when your test app meets the sign-in and project-layout requirements in Run the app with two test users.
If you need Flutter, a game engine, or a remote MCP server, compare the alternatives below. If your app already uses another provider, weigh the integration and migration effort along with the agent tooling. This post compares development tools; it doesn't decide which communications service fits every project.
Compare what each tool covers
We compared each vendor's public agent-skill repository, its plugin manifests, and its official installation instructions as of November 18, 2026. The rows are the jobs the Sendbird Client Plugin handles, so its column says Yes in every row. The other columns show which of those jobs each vendor documents. They are not a ranking of each vendor's product.
"Partly" means the repository covers part of the job, and the note in the cell says which part. "No" means the repository shows that it is missing, for example no plugin manifest or a platform that isn't in its supported list. "Not found" means we didn't find it in the repository, not that the product can't do it. Each cell describes what the vendor documents, not an install or a run we did for this post. Docs that a skill fetches while it runs may cover more. Twilio's repository covers its server APIs and is in public beta.
The table leaves out what other tools do and the Sendbird Client Plugin doesn't. Stream, CometChat, Agora, and TalkJS have Flutter skills. Stream also covers Unity and Unreal, CometChat and TalkJS cover Angular, and TalkJS covers Vue and Svelte. Stream, CometChat, Twilio, and Agora also ship a remote MCP server with their plugins, and the Sendbird Client Plugin doesn't.
Check the code before it is written
Every repository in the table points the agent to official docs before it writes code, and most tell it outright not to invent APIs. Stream's skills say never to answer SDK specifics from training data. CometChat requires every symbol to exist in its catalog for the target UI Kit family and major version, and a compile check before the agent is done. Agora says not to invent fields, flags, or commands that the docs don't list, and Twilio's skills carry explicit CANNOT sections.
These are instructions, and they work when the model follows them. The Sendbird Client Plugin also ships hooks, which Claude Code runs on its own before a file write or a shell command. Codex runs the same hooks once you trust them. In the eight repositories we checked (the five in the table, plus PubNub, Ably, and LiveKit), we found no hooks folder, hooks.json file, or hooks entry in a plugin manifest. That covers bundled hooks only; it doesn't show that their products lack other protections.
These hooks check the agent's work:
- Source check. New Sendbird API code is blocked unless the change cites a source that contains the name: the SDK installed in your project, a plugin file, a page on
sendbird.com/docs, or a file in agithub.com/sendbirdrepository. A file the agent wrote in your project doesn't count. - Credential check. A file write or shell command that contains what looks like a key or token is blocked.
- Pasted-token check. When it detects a token you pasted into the chat (a long value next to a word such as token or secret), a skill or agent call that passes it on is blocked, and an answer that repeats it goes back for revision. This lowers the chance of a leak along those paths; it doesn't guarantee that a token can never be exposed.
- Destructive-command check. A Sendbird CLI
delete,revoke, ordeactivaterun with--yeswaits for your approval. - Build check. While the plugin works on your project, a build or test you didn't ask for is held back.
- Final-answer check. An answer about Sendbird that names APIs without sources goes back for revision.
A hook checks only what it was written to check. The source check reads Sendbird type and method names, not numbers, and it can't tell whether a real API is the right one for the job. The hooks need jq on your PATH. Without it, the credential, source, and destructive-command checks ask you to approve every file write and command, and the other checks don't run.
The plugin's notes carry the values that a name check can't. In our recorded runs on an iOS app, across two requests to show why a message failed to send, Claude Sonnet 5.5 without the plugin put wrong Sendbird error codes into the code in 3 of 8 runs. Nothing in that code would fail to compile, because it compared plain numbers. With the plugin (a development build in most runs), all 8 runs used the SDK's codes from the plugin's iOS error table, and two more runs with a released version did too. These were a few runs, not a benchmark, and we checked the code, not the app at run time.
The README lists every hook, describes each kind of script the agent runs, and lists the network requests the plugin makes, under What runs on your machine.
Add native mobile chat and calls that ring
The plugin has a platform agent for iOS, one for Android, and one for React and React Native. On iOS it installs the SDK with Swift Package Manager, and it uses CocoaPods only when your project already has a Podfile.
On iOS, Android, and React Native, a new incoming 1:1 call arrives only through push, even while the app is open. For an app that already uses Sendbird Calls, the calls-incoming skill checks what makes a call ring: PushKit with CallKit on iOS, FCM with a call notification on Android, and native push on React Native. It lists each gap with the file and line, and changes the code after you say yes.
Among the other repositories, Stream's skills cover ringing on iOS, Android, React Native, and Flutter. CometChat covers it on iOS and React Native and links to its docs for native Android. We didn't find it in the Twilio, Agora, or TalkJS repositories.
Push credentials and turning on Calls are settings of your Sendbird app, not code. The Sendbird CLI plugin can do both from the same session; turning on Calls can't be undone for that app. Signing capabilities and a test on a real device are still your steps.
Run the app with two test users
When you ask for a run check and agree to the build, the run-check skill copies your project to a temporary folder, builds it there, and launches it in a browser, an iOS simulator, or an Android emulator. User A signs in through your app. User B, driven by the plugin's script, opens a channel and sends a message that A's screen must show, and A's reply must reach B. Your app's files aren't changed; results and screenshots go to the plugin's git-ignored .sendbird-client/ folder.
CometChat hands you an acceptance checklist and runs those checks only when you ask, which is also when the Sendbird run check runs. Stream's React and React Native skills tell the agent to test multi-user flows as two users, and TalkJS's skill has the agent send a real message, reply, and reload the chat. The difference is that the Sendbird Client Plugin ships the script that builds, launches, and drives both users, so the steps are the same each time.
| Check | What it establishes |
|---|---|
| Type check or build | The code compiles against the installed SDK |
| Run check passes with every step run | Two users connected and exchanged messages through your Sendbird app, in this build |
| Your own tests | Production sign-in with tokens, push, calls, and your business rules |
The run check has limits. If your app builds its own chat screens with the Chat SDK instead of UIKit, the run check stops after the connection step unless you run it again with the option for those screens, such as --chat-screens. Check which steps ran before you treat a pass as proof that messages were delivered. It signs in with a user ID only, so an app that needs an access token can't be checked this way. It supports React web apps, iOS apps built from an XcodeGen project.yml, Android apps with a Kotlin DSL app/build.gradle.kts, and React Native apps that have their ios/ and android/ folders. The test users, their channel, and the messages stay in your Sendbird app until you delete them.
Read the common questions
Is there an official Sendbird plugin for Claude Code and Codex?
Yes. The Sendbird Client Plugin (sendbird-client) is Sendbird's official plugin for both. It is published in the sendbird/sendbird-plugins marketplace under the Apache-2.0 license, and the same plugin installs in either tool. Its source code is public on GitHub.
Do other chat SDK agent skills block invented APIs?
Other tools give the agent instructions and, in some cases, a symbol catalog or a compile check to reduce invented APIs. The Sendbird Client Plugin also ships hooks that check supported writes before they go through. In the eight public repositories we checked on November 18, 2026, we found no such hooks. The hooks have set coverage and prerequisites, and they don't guarantee that every API call is correct.
Does the Sendbird Client Plugin support Flutter?
Not for setup or code. It writes code for iOS, Android, React, and React Native. For Flutter, follow the Sendbird Chat SDK for Flutter docs.
Does the plugin send my code to Sendbird?
The README states that none of the plugin's network requests sends your project files or conversation to Sendbird. The hooks fetch the public docs pages your code cites, setup installs packages, agents may search public GitHub repositories for an API name through your gh login, and a run check you approve connects test users to your Sendbird app. Claude Code or Codex itself sends your conversation, and the files it reads, to its model provider under its own terms.
Install the plugins and try one request
Install the Sendbird Client Plugin, which writes and checks your app code, and the Sendbird CLI plugin, which connects the agent to your Sendbird app. The hooks need jq on your PATH.
Claude Code
claude plugin marketplace add sendbird/sendbird-plugins
claude plugin install sendbird-client@sendbird
claude plugin install sendbird-cli@sendbird
Codex
codex plugin marketplace add sendbird/sendbird-plugins
codex plugin add sendbird-client@sendbird
codex plugin add sendbird-cli@sendbird
The CLI used by the CLI plugin is available as the Sendbird CLI package on npm; the Client Plugin is installed through your coding agent’s plugin marketplace.
In Codex, when you see "Hooks need review," choose Review hooks and press t. Until you do, the checks don't run.
Then start a new session in your app's folder, sign in and pick a test app with the CLI plugin setup guide, and ask for one thing:
The agent asks before it builds and runs the app. Review the code and the run check results before you merge. If the plugin gets something wrong, open an issue on GitHub.
Start building with Sendbird for free. Add chat to your app with the Client plugin’s Add Chat guide.
Stream, CometChat, Twilio, Agora, TalkJS, PubNub, Ably, and LiveKit are trademarks of their respective owners. This comparison reflects their public agent-skill repositories on November 18, 2026, and they may have changed since.











