OpenCode Go

OpenCode Go provides OpenAI-compatible and Anthropic-compatible endpoints under https://opencode.ai/zen/go/v1. Authentication uses OPENCODE_API_KEY.

Add the key

bash

export OPENCODE_API_KEY="your-opencode-go-api-key"

Configure the Chat Completions models

The six models below use the OpenAI-compatible Chat Completions endpoint:

toml

[model_providers.opencode-go]
base_url = "https://opencode.ai/zen/go/v1"
api_backend = "chat_completions"
env_key = "OPENCODE_API_KEY"
context_window = 200000

[model.kimi-k2-7-code]
model = "kimi-k2.7-code"
model_provider = "opencode-go"
name = "Kimi K2.7 Code"

[model.glm-5-2]
model = "glm-5.2"
model_provider = "opencode-go"
name = "GLM-5.2"

[model.grok-4-5-go]
model = "grok-4.5"
model_provider = "opencode-go"
name = "Grok 4.5 · OpenCode Go"

[model.kimi-k3]
model = "kimi-k3"
model_provider = "opencode-go"
name = "Kimi K3"

[model.deepseek-v4-flash-go]
model = "deepseek-v4-flash"
model_provider = "opencode-go"
name = "DeepSeek V4 Flash · OpenCode Go"

[model.deepseek-v4-pro-go]
model = "deepseek-v4-pro"
model_provider = "opencode-go"
name = "DeepSeek V4 Pro · OpenCode Go"

Current streaming limitation

Direct calls succeeded for all six model IDs, but the current grok-build Chat stream parser rejects OpenCode's trailing inference-cost SSE event. That event has an empty choices array and no normal completion id. The harness must ignore or separately parse this metadata event before these models are fully agent-compatible.

Messages models

For a model listed by OpenCode as Anthropic Messages compatible, configure the protocol and x-api-key authentication on the model:

toml

[model.opencode-messages]
model = "minimax-m3"
base_url = "https://opencode.ai/zen/go/v1"
api_backend = "messages"
auth_scheme = "x_api_key"
env_key = "OPENCODE_API_KEY"
context_window = 200000
extra_headers = { "anthropic-version" = "2023-06-01" }

This path has completed an end-to-end grok-build smoke test with minimax-m3. Confirm the current model catalog before using a different ID.

Inspect the current catalog

This prints model IDs without printing the key:

bash

curl --fail --silent \
  -H "Authorization: Bearer ${OPENCODE_API_KEY}" \
  https://opencode.ai/zen/go/v1/models

Provider ZDR statement

OpenCode states that Go requests have zero retention and are not used for training. Treat that as the provider's policy statement and verify the current terms for your account before approving sensitive workloads.

Official references:

OpenCode Go ​

Add the key ​

Configure the Chat Completions models ​

Messages models ​

Inspect the current catalog ​