Deploy iCode in your network
iCode is one service container next to a model server. A first install takes an afternoon; the deployment kit includes the full installation guide for production sites.
What you need
Service host
A Linux server with a container runtime and compose. A pilot for up to 10 developers needs 4 vCPU, 8 GB memory and 50 GB of SSD.
A model
A GPU host for the bundled model server, or the OpenAI-compatible AI gateway you already run. See Models.
Sign-in
Your identity provider (any OpenID Connect provider) and the groups that map to iCode roles. A licence file from IOTA.
Quick start (private network with internet access)
-
Unpack the deployment kit and load the image
tar xzf icode-deploy-kit.tar.gz && cd icode-deploy-kit docker load < ../icode-service-image.tar.gz
-
Fill in the settings
cp stack.env.example stack.env # set MODEL_GATEWAY_API_KEY, OIDC_ISSUER and SESSION_SIGNING_SECRET mkdir -p etc/licenses data logs repos cp /path/to/license-public.pem etc/ && cp /path/to/*.lic etc/licenses/
-
Start it
docker compose --env-file ./stack.env up -d curl -s http://127.0.0.1:8787/healthz # {"ok":true,"version":"0.23.1",...} -
Sign in from a client
export ICLI_SERVICE_URL=http://<service-host>:8787 icli login && icli status # shows your role, tenant, licence and model
Air-gapped install
For networks with no internet at all, everything comes in one file: the service image, the deployment kit, the three clients, the guide and a checksum list. Model weights travel separately because of their size.
-
Bring the bundle across on approved media
Copy
icode-airgapped.tarto the service host. The licence file and the licence public key come from IOTA separately, never inside the bundle. -
Check it before unpacking anything
tar xf icode-airgapped.tar && cd icode-0.23.1 gpg --import iota-release-key.asc # IOTA's release key, delivered out of band gpg --verify SHA256SUMS.sig SHA256SUMS sha256sum -c SHA256SUMS # every line must say OK
-
Load the image and start
docker load < icode-service-image.tar.gz tar xzf icode-deploy-kit.tar.gz && cd icode-deploy-kit cp stack.env.example stack.env # fill in the REQUIRED lines docker compose --env-file ./stack.env up -d
-
Hand the clients to developers
The bundle's
clients/folder has the VS Code extension, the JetBrains plugin and icli. None of them download anything when installed.
Keep the previous release folder until the new one passes your checks. Rolling back is a restart against the old folder.
Models
iCode talks to any OpenAI-compatible endpoint, so you choose the model and where it runs.
- The bundled model server. The compose file includes one, started on a GPU host with weights you copy in. It never downloads anything at run time.
- Your existing AI gateway. Point the service at the gateway's base URL and key, and start only the iCode service.
Once a model is connected, iCode measures it and prints the settings that suit it, so a smaller or larger model works without guesswork. Run the check again whenever the model changes.
Clients
code --install-extension icode-vscode.vsix # VS Code npm install -g ./icli.tgz # terminal, Node.js 20 or later
JetBrains IDEs: Settings, Plugins, the gear icon, Install Plugin from Disk, then icode-intellij.zip.
In each client, set the iCode service address (for example https://icode.your-company.local) and sign in with your company account. Your custom certificate authority is picked up from the operating system.
Updates
Every release has one version shared by the service and all clients. To update, load the new image, set ICODE_IMAGE in stack.env to the new version and run docker compose up -d again. Clients update by installing the new file over the old one; icli version --check tells developers when the service is newer than their client.