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:

Window	Share of monthly allowance
5 hours	20%
Week	50%
Month	100%

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

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

Custom model

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

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

Protocol matrix

Protocol	Auth	Session	UA
chat		required	required
anthropic	+	required	required
responses		required	required

Base URL default: .

Custom model form (0.4.0)

When adding a custom model you can now set:

Field	Purpose
Model ID	Upstream model id (required)
Display Name	Label in the model list (optional; defaults to Model ID)
Thinking / Agent Thought	Default on. Adds + thinking parameters
Vision / Audio / Video / Document	Multimodal file support
Structured output	Adds /
API Protocol	/ /
Function calling / context / max tokens	As before

Predefined models added in 0.5.0

Model	Protocol	Notes
Claude Haiku 5.5 ()	anthropic	Messages-only (). Base $0.10 / $0.50 (≤ 100K input); long-context tier > 100K $0.50 / $2.50. Context 1M, max output 128K. Prompts are not used for training; 30-day data retention. Listed in the official catalog and on the provider page, but as of 2026-10-08 the gateway did not return this id yet (may require billing activation). Not smoke-tested.
LongCat 2.5 Preview Free ()	chat	Free for a limited time (Unlimited during the promo) — may be removed anytime. Context 1M, max output 131,072. Do not rely on it for production.
Space Bunny ()	chat	Renamed from (see removal table below) and now paid: $0.15 / $0.60, cache read $0.03, $30 monthly allowance.

Predefined models added in 0.3.0

Model	Protocol	Notes
Grok 4.7 ()	responses	Same Responses-only path as Grok 4.6. May require outbound proxy from some regions (e.g. CN).
MiMo-V2.6-Flash ()	chat	Aligns with MiMo-V2.5 multimodal flags (vision / video / audio)
MiMo-V2.6-Pro ()	chat	Aligns with MiMo-V2.5-Pro
GPT 6 Luna ()	responses	Documented Responses-only. Region-restricted in some territories.
Space Bunny Free ()	chat	Free for a limited time — may be removed anytime. Do not rely on it for production.

Predefined models added in 0.2.0

Model	Protocol	Notes
MiniMax M2.7 ()	anthropic	oa-compat returns 500; works
Grok 4.6 ()	responses	Documented Responses-only. May require outbound proxy from some regions (e.g. CN).
GPT 5.6 Luna ()	responses	Documented Responses-only; region-restricted in some territories (often needs proxy). Plugin strips / for this model.
Muse Spark 1.3 Contributor	responses	Region-limited (Meta geographic policy); contributor tier may use prompts for training
Muse Spark 1.2 Contributor	responses	Same as above

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

Model	Reason
Space Bunny Free ()	Upstream renamed the model id to — the gateway now only returns , so no longer resolves. The model also became paid ($0.15 / $0.60). If you had added it as a custom model, re-create it with the id .

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

Model	Reason
Union Alpha Free ()	No longer listed in the OpenCode Go catalog (was a limited-time free trial). Custom models can still target the id if the gateway accepts it.
MiniMax M2.5 ()	No longer in the Go "current model list" / usage tables. Removed from predefined models.

Deprecation notices

Model	Offline at (UTC+8)	Migration
	2026-10-21 10:00	Switch to
	2026-10-21 10:00	Switch to

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

Model	Behavior
	Gateway only accepts and ; the plugin overrides other values.
	Upstream rejects and ; the plugin strips them.

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):

Model	Cache read	Cache write
	$0.01	$0.125
/ /	$0.003	—
	$0.022	—
/ /	$0.26	—
	$0.03	—
	$0.02	$0.25
	$0.01	$0.125
/	$0.50	—
	$0.035	—
	$0.042	—
	$0.16	—
	$0.19	—
	$0.30	—
	$0.006	—
	Free	Free
/	$0.0028	—
/	$0.003625	—
	$0.06	$0.375
	$0.06	—
/	$0.002	—
	$0.05	$0.625
	$0.04	$0.50
	$0.50	$3.125
	$0.016	$0.20
	$0.25	$2.50
	$0.03	—

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

Model	Threshold	Tier input / output	Tier cache read / write
	> 100,000	$0.50 / $2.50	$0.05 / $0.625
	> 272,000	$0.40 / $1.80	$0.04 / $0.50
	> 272,000	$0.20 / $0.75	$0.02 / $0.25
/	> 200,000	$4.00 / $12.00	$1.00 / —
	> 512,000	$0.60 / $2.40	$0.12 / —
	> 256,000	$2.00 / $6.00	$0.20 / $2.50
	> 256,000	$1.20 / $4.80	$0.12 / $1.50

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.

Model	Off-Peak input / output	Peak input / output
	$0.15 / $0.60	$0.30 / $1.20
	$0.66 / $1.98	$1.32 / $3.96
	$0.15 / $0.60	$0.30 / $1.20
	$0.15 / $0.60	$0.30 / $1.20

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:

Model	Status
glm-5.3-flash / glm-5.x	OK (stream + non-stream)
glm-5.1 / glm-5.2	OK
mimo-v2.6-flash / mimo-v2.6-pro	Not smoke-tested yet (added in 0.3.0)
mimo-v2.5 / mimo-v2.5-pro	OK
kimi-k2.6 / kimi-k2.7-code / kimi-k3	OK
qwen3.6-plus / qwen3.7-plus / qwen3.7-max / qwen3.8-flash / qwen3.8-max	OK
minimax-m3	OK (chat)
minimax-m2.7	OK (anthropic )
deepseek-v4-pro / v4-flash / v4.1-flash / flash-vision-exp	OK
longcat-2.0 / hy3 / hy4-preview	OK
grok-4.7 via responses	Not smoke-tested yet (added in 0.3.0)
grok-4.6 via responses	OK with outbound proxy
gpt-5.6-luna via responses	OK with outbound proxy
muse-spark-1.3 via responses	HTTP 200 (content quality may vary; region-limited)
claude-haiku-5-5 via anthropic	Not smoke-tested yet (added in 0.5.0; gateway did not list it as of 2026-10-08)
longcat-2.5-preview-free	Not smoke-tested yet (added in 0.5.0)
space-bunny	Not smoke-tested yet (added in 0.5.0; renamed from )

Development / debug

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

Local unit tests (no network):

Live smoke (needs ):

Package:

Disclaimer

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

OpenCode Go

OpenCode Go

Features

Plans & usage limits

Setup

Custom model

Protocol matrix

Custom model form (0.4.0)

Predefined models added in 0.5.0

Predefined models added in 0.3.0

Predefined models added in 0.2.0

Predefined models removed in 0.5.0

Models kept outside the official catalog

Predefined models removed in 0.3.0

Deprecation notices

Model parameter constraints

Cache & tier pricing (reference)

Multimodal flags

Development / debug

Links

Disclaimer