opencode-ollama.md

OpenCode client for Ollama

Retired. Ollama no longer runs anywhere: the Tesla P40 cards moved to adler-black-u2, and the VM keeps only the GPU-less recorder profile. OpenCode's local model is black-qwen (krotname/VpnOps, ops/black-qwen/). This page documents the gpu_analytics profile only.

The files in opencode/ configure OpenCode to use the authenticated OpenAI-compatible Ollama endpoint without committing the URL, credentials, or local CA certificate.

The configuration was runtime-tested with OpenCode 1.17.11 and Ollama 0.30.8 on 2026-07-12. All four declared models returned a response. Qwen3 also completed a native glob tool call and then returned the requested JSON.

Prepare the model tags

The OpenCode configuration expects four exact local tags. The general IaC deployment pulls only its primary Ollama model, so create the client-specific tags before installing this configuration. Run these commands on the Ollama VM from a checkout or transferred copy that includes opencode/models/:

ollama_local() {
  sudo -u ollama -H env OLLAMA_HOST=http://127.0.0.1:11435 ollama "$@"
}

ollama_local pull huihui_ai/gpt-oss-abliterated:20b
ollama_local pull huihui_ai/qwen3-abliterated:8b
ollama_local pull huihui_ai/nemotron-v1-abliterated:8b-llama-3.1-nano
ollama_local pull huihui_ai/mistral-small-abliterated:24b

ollama_local create gpt-oss-uncensored:16k \
  -f opencode/models/gpt-oss-uncensored-16k.Modelfile
ollama_local create qwen3-uncensored:16k \
  -f opencode/models/qwen3-uncensored-16k.Modelfile
ollama_local create nemotron-uncensored:16k \
  -f opencode/models/nemotron-uncensored-16k.Modelfile
ollama_local create mistral-adler:16k \
  -f opencode/models/mistral-adler-16k.Modelfile

The host-local 11435 endpoint is created by this repository's nginx deployment. Do not use the default 127.0.0.1:11434 on this VM.

The pulls are large. Do not repeat them during every client install. Existing blobs are reused when the tags are already present.

Install on Windows

Copy the example and global rules into the OpenCode configuration directory:

$configDir = Join-Path $HOME ".config\opencode"
New-Item -ItemType Directory -Path $configDir -Force | Out-Null
Copy-Item .\opencode\opencode.example.json (Join-Path $configDir "opencode.json")
Copy-Item .\opencode\AGENTS.md (Join-Path $configDir "AGENTS.md")

Set the endpoint without a trailing slash because the configuration appends /v1. Store the Base64 value of username:password, not the already prefixed HTTP header:

$basicUser = Read-Host "Ollama basic-auth user"
$basicPassword = Read-Host "Ollama basic-auth password" -AsSecureString
$credential = [pscredential]::new($basicUser, $basicPassword)
$pair = "{0}:{1}" -f $credential.UserName, $credential.GetNetworkCredential().Password
$basicB64 = [Convert]::ToBase64String([Text.Encoding]::UTF8.GetBytes($pair))

[Environment]::SetEnvironmentVariable(
    "ADLER_OLLAMA_URL",
    "https://ollama-host.example:11443",
    "User"
)
[Environment]::SetEnvironmentVariable("ADLER_BASIC_B64", $basicB64, "User")
[Environment]::SetEnvironmentVariable(
    "NODE_EXTRA_CA_CERTS",
    "C:\path\to\local-root-ca.pem",
    "User"
)

Open a new terminal after changing user environment variables. For immediate validation in the current PowerShell process, reload them explicitly:

$env:ADLER_OLLAMA_URL = [Environment]::GetEnvironmentVariable("ADLER_OLLAMA_URL", "User")
$env:ADLER_BASIC_B64 = [Environment]::GetEnvironmentVariable("ADLER_BASIC_B64", "User")
$env:NODE_EXTRA_CA_CERTS = [Environment]::GetEnvironmentVariable("NODE_EXTRA_CA_CERTS", "User")

Test-Path -LiteralPath $env:NODE_EXTRA_CA_CERTS
$resolved = opencode debug config | ConvertFrom-Json
[pscustomobject]@{
    ProviderPresent = $null -ne $resolved.provider.adler
    EndpointPresent = -not [string]::IsNullOrWhiteSpace($resolved.provider.adler.options.baseURL)
    AuthPresent = -not [string]::IsNullOrWhiteSpace(
        $resolved.provider.adler.options.headers.Authorization
    )
    Model = $resolved.model
    SmallModel = $resolved.small_model
}

$headers = @{ Authorization = "Basic $env:ADLER_BASIC_B64" }
$available = @(
    (Invoke-RestMethod -Uri "$env:ADLER_OLLAMA_URL/v1/models" -Headers $headers).data.id
)
$expected = @(
    "gpt-oss-uncensored:16k",
    "qwen3-uncensored:16k",
    "mistral-adler:16k",
    "nemotron-uncensored:16k"
)
$missing = @($expected | Where-Object { $_ -notin $available })
if ($missing.Count -gt 0) {
    throw "Missing Ollama model tags: $($missing -join ', ')"
}

opencode run --model adler/qwen3-uncensored:16k "Reply with exactly: OK"

Do not print or log the full $resolved object: it contains the resolved Authorization header.

OLLAMA_HOST is not used by this OpenCode provider. Connection settings come from ADLER_OLLAMA_URL and ADLER_BASIC_B64.

Models

OpenCode model	Purpose
adler/gpt-oss-uncensored:16k	Default reasoning model
adler/qwen3-uncensored:16k	Small model, titles, and native tool calls
adler/mistral-adler:16k	General 24B model
adler/nemotron-uncensored:16k	Compact non-reasoning model

The first request after a cold model load can take several minutes. The client timeout is therefore set to 15 minutes.

Secret hygiene

Do not commit a real ADLER_OLLAMA_URL, ADLER_BASIC_B64, CA certificate, password, or a resolved Authorization header. The committed configuration contains environment placeholders only.