OpenCode

opencode-ai (v1.14.33+). Active upstream is github.com/anomalyco/opencode (sst/opencode redirects there).

Install

npm install -g opencode-ai

Auth (BYOK across 75+ providers)

OpenCode reads provider credentials from env vars based on which provider/model you pick. Crewship’s apiKeyEnvVarsForAdapter("OPENCODE") accepts:

Env var	Provider
`ANTHROPIC_API_KEY`	Anthropic
`OPENAI_API_KEY`	OpenAI
`GOOGLE_API_KEY` / `GEMINI_API_KEY`	Google
`OPENROUTER_API_KEY`	OpenRouter
`XAI_API_KEY`	xAI / Grok
`GROQ_API_KEY`	Groq
`DEEPSEEK_API_KEY`	DeepSeek

Add any of the above as a SECRET-type credential and assign to your OpenCode agent.

Models

OpenCode requires provider/model namespaced strings (NOT bare IDs). Curated picker entries — opencode/ rows bill through a Zen account, opencode-go/ through a Go subscription, zai-coding-plan/ through a GLM Coding Plan:

opencode/claude-sonnet-5 | opencode/gpt-5.5 | opencode/gemini-3.1-pro | opencode/kimi-k3
opencode-go/kimi-k3 | opencode-go/glm-5.3-flash | opencode-go/minimax-m3 | opencode-go/gpt-5.6-luna
zai-coding-plan/glm-5.3 | zai-coding-plan/glm-5.3-flash | zai-coding-plan/glm-5.3-highspeed | zai-coding-plan/glm-5.2
anthropic/claude-fable-5 | anthropic/claude-opus-4-8 | anthropic/claude-sonnet-5 | anthropic/claude-opus-4-7 | anthropic/claude-sonnet-4-6 | anthropic/claude-haiku-4-5
openai/gpt-5.5 | openai/gpt-5.4 | openai/gpt-5.4-mini | openai/gpt-5.4-nano | openai/gpt-5.3-codex | openai/o3 | openai/o3-pro
google/gemini-3.1-pro-preview | google/gemini-3-flash-preview | google/gemini-2.5-pro | google/gemini-2.5-flash
xai/grok-4.1-fast | xai/grok-4
deepseek/deepseek-v4-flash | deepseek/deepseek-v3.2 | deepseek/deepseek-r1
groq/llama-3.3-70b-versatile | groq/qwen-2.5-coder-32b
moonshotai/kimi-k2.6 | zai/glm-5.1 | minimax/minimax-m2.7
openrouter/anthropic/claude-sonnet-4-6 | openrouter/openai/gpt-5.5
ollama/qwen2.5-coder:7b | ollama/qwen3-coder:30b | ollama/devstral:24b

Command shape

opencode run --format json --model <provider/model> -- <user message>

--format json (NOT --output-format) emits JSONL events.

MCP

Configured at /output/<slug>/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "fs": {
      "type": "local",
      "command": ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/work"],
      "environment": {"FOO": "{env:FOO}"},
      "enabled": true
    },
    "linear": {
      "type": "remote",
      "url": "https://mcp.linear.app/sse",
      "headers": {"Authorization": "Bearer {env:LINEAR_TOKEN}"},
      "enabled": true
    }
  }
}

OpenCode-specific:

Top key is mcp (NOT mcpServers)
type: local (stdio) or remote (HTTP/SSE)
command is an ARRAY (binary + args together)
env field is environment (NOT env)
Each entry carries "enabled": true
Env-var syntax is {env:VAR} (NOT ${VAR}) — Crewship translates standard ${VAR} form on write

Crewship auto-injects two sidecar-hosted servers under mcp (injectMemoryMCP + injectRoutinesMCP): crewship-memory and crewship-routines, both type: remote pointing at http://127.0.0.1:9119/mcp/memory/<slug> and .../mcp/routines. A user-declared server of the same name wins.

Memory

AGENTS.md is OpenCode’s auto-discovered system instructions file. Crewship writes it as part of the canonical memory set. For turn-1 parity (OpenCode discovers AGENTS.md on the second invocation, not the first), the orchestrator also prepends the preamble inline with [SYSTEM]/[USER] delimiters on the first turn — turn-2+ uses the file.

Output stream

Flat JSONL per packages/opencode/src/cli/cmd/run.ts:

{type: "text", sessionID: "...", part: {id, text}}
{type: "reasoning", part: {text}}
{type: "tool_use", part: {id, tool, state: {status, input, output}}}
{type: "step_start", part: {}}
{type: "step_finish", part: {tokens, cost, providerID, modelID}}
{type: "error", error: "..."}

Note type IS the part type at top level — NOT a nested message.part.updated envelope. step_finish (snake_case, NOT hyphen).

Troubleshooting

Bare model rejected — use provider/model form (anthropic/claude-sonnet-4-6 not claude-sonnet-4-6). MCP server uses literal ${VAR} token — Cursor/Claude ${VAR} syntax incompatible; ensure Crewship’s writer translated to {env:VAR}. Check by cat /output/<slug>/opencode.json.

OpenCode Go and Zen

In Credentials → Add provider, choose OpenCode Go or OpenCode Zen, follow the console link and paste an API key. Name the account and assign it to the crew or agent that will use it. Then select the matching model provider in the agent’s model settings; the runner is OpenCode.

Go models use opencode-go/<model> and the Go subscription endpoint.
Zen models use opencode/<model> and the metered Zen endpoint.

These are separate accounts in Crewship even if you reuse the same upstream key. The selected key stays in the sidecar; OpenCode receives dummy auth and keeps its native per-model protocol. Crewship’s credential slots are OPENCODE_API_KEY for Zen (mapped to opencode in OpenCode’s auth file) and OPENCODE_GO_API_KEY for Go (mapped to opencode-go) — upstream environment-variable conventions these are not. Go connects using an API key, but has subscription usage limits. Its optional Use balance setting can consume Zen credit after those limits; manage it in the OpenCode console. Crewship does not infer remaining subscription quota or balance-overage charges from token counts. Zen usage estimates use Zen’s model prices. The console is authoritative for both products’ actual billing, limits and model access. Model suggestions are a curated subset; use a custom native model ID for other models your OpenCode version supports. Transport verified with OpenCode 1.18.31; update older crew installations if the provider or model is unknown. See the official Go and Zen documentation.

Z.AI GLM Coding Plan

In Credentials → Add provider, choose Z.AI Coding Plan, subscribe to the GLM Coding Plan at Z.AI, create an API key and paste it below. Name the account and assign it to the crew or agent that will use it. The runner is OpenCode and models use the zai-coding-plan/ prefix (for example zai-coding-plan/glm-5.3). This is a subscription product, not metered Z.AI API credit: a different endpoint (api.z.ai/api/coding/paas/v4), a different key and its own usage limits. The two are separate accounts in Crewship even if the same vendor issued both keys, and one is never used as a fallback payer for the other. A metered Z.AI key stored under Z.AI does not authenticate Coding Plan models — the run fails closed instead of switching products. The selected key stays in the sidecar; OpenCode receives dummy auth and keeps its native openai-compatible protocol. Crewship’s credential slot is ZAI_CODING_PLAN_API_KEY, mapped to zai-coding-plan in OpenCode’s auth file; upstream itself reads ZHIPU_API_KEY for every z.ai/zhipu product, so that shared name is not used as an account identity here. Crewship does not infer remaining plan quota or overage pricing from token counts; the Z.AI console is authoritative for limits and billing. Model suggestions are the catalog entries of the deployed OpenCode version (verified against 1.18.31); use a custom native model ID for others your plan supports. Zhipu/BigModel regional variants are separate products and are not configured by this connection.

Get Started

Guides

Security

Configuration

Manifest Reference

Install

Auth (BYOK across 75+ providers)

Models

Command shape

MCP

Memory

Output stream

Troubleshooting

OpenCode Go and Zen

Z.AI GLM Coding Plan

​Install

​Auth (BYOK across 75+ providers)

​Models

​Command shape

​MCP

​Memory

​Output stream

​Troubleshooting

​OpenCode Go and Zen

​Z.AI GLM Coding Plan

Install

Auth (BYOK across 75+ providers)

Models

Command shape

MCP

Memory

Output stream

Troubleshooting

OpenCode Go and Zen

Z.AI GLM Coding Plan