OpenCode Go

Dify model provider plugin for OpenCode Go.

OpenCode Go is a subscription gateway for curated open coding models with two tiers: Go ($10/month) and Go Plus ($40/month) — identical per-token pricing, different monthly allowances. This plugin exposes those models as a Dify LLM provider.

Features

• Predefined models from the OpenCode Go catalog (GLM, Kimi, DeepSeek, MiMo, MiniMax, Qwen, LongCat (incl. LongCat 2.5), Claude, Hy, Grok, GPT Luna, Muse Spark, Space Bunny)
• Customizable model support with an API Protocol selector ( / / )
• Three upstream protocols in one provider:
  • Chat Completions ( + ) — default for most models
  • Anthropic Messages ( + ) — models that only expose (e.g. , )
  • OpenAI Responses ( + ) — models that only expose (e.g. , , , , )
• Sends OpenCode-required headers on all three paths:
  • : (not a generic SDK name)
  • : stable id for routing / prompt-cache affinity
• Upstream quirks handled automatically:
  • — forces / (gateway only accepts these)
  • — strips / (upstream rejects them)
  • Transient / / are retried with backoff; empty SSE / raw error JSON surface as Dify (never silent empty text)
  • SSE is always decoded as UTF-8 (OpenCode often omits charset)
  • Outbound requests honor the process / system proxy ()
• Session isolation (recommended): enable the LLM-node model parameter and keep the default JSON. Dify resolves before invoke; the plugin then picks:
  • Chatflow / chat apps: conversation id → one session per conversation
  • Workflow apps: (via the internal helper header) → one session per run, shared by LLM nodes in that run

• Fallbacks when is absent or leaves session empty:
  1. Provider credential (optional static override)
  2. Plugin Session when Dify provides it
  3. Per-invoke isolation (RPC session id or a random UUID) — never sticky on Dify user id
• Unresolved Dify templates () are never sent as session values.
• The internal helper header is never forwarded upstream.

Plans & usage limits

Two subscription tiers — Go ($10/month) and Go Plus ($40/month) — share the same per-token pricing; only the monthly allowance differs. Limits are enforced over rolling windows:

When a limit is exceeded, requests can fall back to pay-as-you-go billing if the "Use balance" option is enabled in your OpenCode Console and a balance is available.

Setup

1. Subscribe to OpenCode Go at opencode.ai/auth and copy your API key.
2. Install this plugin in Dify (Marketplace / plugin package / debug remote).
3. Open Settings → Model Providers → OpenCode Go, paste the API key, save.
4. Select an OpenCode Go model in your app.

Custom model

If OpenCode adds a new model before this plugin is updated:

1. Add a custom model under OpenCode Go.
2. Model ID = model id from the Go docs (e.g. ).
3. Display Name (optional) = label shown in the model list. Defaults to the Model ID.
4. Set API Protocol to match the model’s endpoint:
   • (default) →
   • → (, , other Messages-only ids)
   • → (, , , )
5. Configure capability toggles as needed:
   • Thinking / Agent Thought (default on) — exposes thinking parameters (, , )
   • Vision / Audio / Video / Document — multimodal file support
   • Structured output — exposes /
6. Optionally set context size / max tokens / function calling.

Protocol matrix

Base URL default: .

Custom model form (0.4.0)

When adding a custom model you can now set:

Predefined models added in 0.5.0

Predefined models added in 0.3.0

Predefined models added in 0.2.0

Proxy note: Responses-line models are frequently blocked or geo-restricted.
From mainland China, set / (or enable system proxy)
before starting the plugin / Dify runtime so outbound HTTPS can leave the region.
The plugin honors the process / system proxy environment ().

Qwen and MiniMax M3 remain on Chat Completions (oa-compat) even though OpenCode docs list as preferred — oa-compat is verified 200 and avoids regressions. MiniMax M2.7 is the exception (chat 500 → routed to anthropic).

Predefined models removed in 0.5.0

Models kept outside the official catalog

, and are no longer advertised in the official OpenCode Go catalog, but the gateway still serves them and they keep working. The plugin retains them as predefined models for existing workloads; they may be removed in a future release if the gateway drops them.

Predefined models removed in 0.3.0

Deprecation notices

Both models remain fully usable (not disabled) — only the name and description carry an offline notice so existing workloads keep running. Please migrate before the offline date.

Model parameter constraints

Cache & tier pricing (reference)

Dify's only stores base input / output unit prices (USD per 1M tokens). The OpenCode Go catalog also lists cache read/write prices and long-context tiers; those are not enforced by the plugin UI or billing fields, but are listed here for cost planning. Source: OpenCode Go provider catalog (2026-10 snapshot).

Cache pricing (when the gateway applies prompt caching):

Long-context tiers (higher unit price once input exceeds the threshold; YAML always records the base tier):

Models not listed above have no cache or tier rows in the catalog (flat base price only).

Peak / off-peak pricing (DeepSeek V4 line): these models use time-of-day pricing. Peak = Monday–Friday 01:00–04:00 and 06:00–10:00 UTC; every other hour (including weekends) is Off-Peak.

Privacy notes (2026-10): Muse Spark Contributor models use your prompts to train Meta models and are not zero-data-retention (ZDR). DeepSeek's ZDR agreement is renewed through 2026-10-31. Grok / GPT / Claude models retain data for 30 days; all other models retain 0 days.

Multimodal flags

Feature flags (vision / video / document / audio) follow the official model capabilities, because OpenCode Go proxies those upstream APIs. Not every modality is re-tested on the gateway for every model.

Local smoke matrix (2026-09-17, 0.2.0 + live-debug fixes). Vision for all vision-flagged models was verified in a Dify workflow by the user:

Development / debug

Copy to and set your Dify debug key from Plugins → debug.

Local unit tests (no network):

Live smoke (needs ):

Package:

Links

• OpenCode Go docs: https://opencode.ai/v2/docs/console/go
• Models list API:
• Auth / API keys: https://opencode.ai/auth
• Dify plugin docs: https://docs.dify.ai/develop-plugin/dev-guides-and-walkthroughs/creating-new-model-provider

Disclaimer

This is an unofficial community plugin. It is not affiliated with OpenCode / Anomaly or Dify.