Codeg builds three Rust binaries, and its MCP companion must sit next to the server
Collaborative multi-agent AI coding workspace: aggregate sessions from Claude Code, Codex, OpenCode, Pi, Grok Build, etc. Desktop app, self-hosted server, or Docker.
At a glance
- What is it?
- Codeg is a workspace for running many coding agents in one interface, offered as a desktop app, a server or a container. Its build is a three-stage image producing three Rust binaries, with a lookup that depends on the server and its MCP companion being neighbours, and its container publishes ten ports so agents can start dev servers inside it.
- Who is it for?
- Codeg is worth a look if you juggle more than one coding agent and want their sessions in one place rather than four terminals, and the session portability plus the delegation syntax are the two features that justify it. Three things to check before you rely on it.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Rust, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- xintaofei/codeg4,484 ★ · +1,069
| Repository | 2026-09-12 | 2026-11-15 |
|---|---|---|
| xintaofei/codeg | 3.4k | 4.5k |
1,069 Stars in the last 64 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
One release, cut three times in eight hours
The three most recent releases are v0.41.2 on 2026-10-29, v0.41.3 on 2026-10-30 and v0.41.4 on 2026-11-02, each titled with the tool's name and version, and package.json carries version 0.41.4. The branch was last pushed on 2026-11-02, the same day as the latest release.
Ten extra ports, or a wildcard hostname instead
The compose file publishes one port for the application and then a second range of ten, with a comment explaining why: agents start development servers inside the container, and the workbench needs to show those addresses to you. The same comment offers an alternative that avoids the range entirely, where a dev server is named by hostname so that a request for a subdomain reaches the right container port through the one published port. That alternative needs a wildcard DNS record and, behind a reverse proxy, a wildcard virtual host that passes the Host header through, with an example configuration given inline. There is also an automatic mode that inserts one label in front of whatever host the workbench itself was opened at, and an option to disable the bridge entirely.
In-place upgrades die with the container
The longest comment in the compose file is about upgrades, and it is a warning rather than a feature description. The in-place upgrade rewrites the binaries and the web assets inside the container's writable layer rather than in the image. The data volume persists, so your sessions and settings survive, but the upgraded binaries do not: recreate the container, whether by forcing a recreate or by pulling a newer image, and you start from the image's binaries again. For a self-hosted install that means an upgrade is not durable, and the only durable path is a new image. The environment block around it is also worth a look, since the host is set to all interfaces, the port is fixed, and the access token variable defaults to an empty value.
The runtime image pins an ICU version and explains the failure
The final stage of the image installs a long list of packages, and the explanation for the last one is three lines long. A companion binary for office document handling ships with its own embedded runtime, which needs the system internationalisation library at startup. The base image for this stage bundles its own copy statically and therefore does not install the system library, so without the explicit package every invocation of that binary aborts with a message about not finding a valid package, which breaks both skill synchronisation and office file previews in the server and container modes. The comment then pins the version, 72, to Debian bookworm, the distribution the base image uses. It is the most useful comment in the build.
Two binaries, two feature flags, one neighbour lookup
The backend stage compiles three executables in release mode, each with its own feature flag and with default features switched off: the server, codeg-cli, and a companion described in a comment as the model context protocol companion that the runtime injects into each session, and the comment cites the module that does the injecting. The important detail is the constraint: the companion must ship next to the server binary, because the code that finds it looks for it as an executable sibling. That is a packaging invariant, so anything that relocates one binary without the other will break in a way that is not obvious from the compose file.
The postinstall copies a dependency's files into the public directory
The web application's manifest has a postinstall hook written as an inline script rather than a named file. It copies a directory out of the installed editor component's package into the project's public directory, recursively and overwriting, and then rewrites one copied file to strip a source map comment from its final line. That is a build step that reaches into another package's internals and edits them in place inside your own tree, which means a dependency bump can silently change what the copy contains. The rest of the manifest is more conventional: a development server using Turbopack, a production build, tests in four flavours, and two server commands that shell into the desktop project's directory to run the Rust build with the right feature flag.
Every sponsor is a model access service with a referral link
The sponsorship block is worth reading for what it says about the project. Each entry is a commercial service selling access to AI model APIs, several describing themselves as relay or aggregation platforms, and each link carries a referral or campaign parameter. Several offer trial credits on sign-up through that link, with amounts stated in local currency, and one advertises officially proxied access to overseas models alongside support for two of the agents this workspace aggregates. None of that is hidden, and sponsorship blocks routinely carry affiliate links. It does mean the funding is entirely aligned with routing model traffic through third parties, which is the same subject the project exists to aggregate, and a reader deciding where to point their agent credentials should notice that.
Editorial conclusion
Codeg is worth a look if you juggle more than one coding agent and want their sessions in one place rather than four terminals, and the session portability plus the delegation syntax are the two features that justify it. Three things to check before you rely on it. The container publishes a range of ten extra ports for dev servers that agents start, and its host pattern alternative needs wildcard DNS and a reverse proxy, so decide which of those you actually want to expose. In-place upgrades write into the container's writable layer, so an upgrade is lost when you recreate it. And the sponsorship block is entirely made up of model access resellers with referral links, which is a fact worth holding in mind about where the project's funding comes from.
Frequently asked questions
What is Codeg?
A multi-agent coding workspace: it runs several coding agents in one interface, renders each as the same structured conversation rather than a terminal, imports and resumes the session history that agents keep on disk, and lets one agent delegate to another or queue unattended tasks in separate worktrees.
Which coding agents does Codeg support?
The README says twenty agents are built in and that any agent speaking the agent client protocol can join. The project description names Claude Code, Codex, OpenCode, Pi and Grok Build among the ones it aggregates.
How do I run Codeg?
As a desktop app, a self-hosted server or a container. The compose file builds from the repository root, publishes one port for the application plus a ten-port range for development servers started by agents inside it, and keeps data in a named volume. A pre-built image is offered as a commented alternative.
Can I use Codeg on a phone?
The feature table lists iPhone, iPad and Android alongside the desktop, server and container options, and it also names three chat platforms as surfaces.
What is the MCP companion binary Codeg ships?
A second Rust binary built beside the server in the same build step with its own feature flag. The Dockerfile comment describes it as the standard input companion the runtime injects per session, and says it must sit next to the server binary because the lookup finds it as an executable sibling.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/xintaofei-codeg)Star history chart for your README
[](https://hysenlabs.com/projects/xintaofei-codeg)Was this useful?