OpenCode Cowork Proxy Worker

This project lets Claude use OpenCode Go models, and some OpenCode Zen models.

Claude normally speaks the Anthropic API format. OpenCode Go mostly speaks OpenAI-compatible API format. This small Cloudflare Worker sits in the middle and translates between them.

I covered how to set this up in Claude in How to Use Claude Code for Free with OpenCode.

Image / Vision Support

When you attach an image in Claude Code and send it through this proxy, the request is automatically routed to MiMo-V2.5 Free (mimo-v2.5-free) — a free vision-capable model available on OpenCode Zen. The proxy detects image blocks in your request, translates them to OpenAI's image format, and overrides the model to MiMo-V2.5 Free so the model can actually see the image.

Hybrid routing: when your base URL uses the /go prefix, the proxy stays on Go for text requests (deepseek-v4-flash on https://opencode.ai/zen/go/v1) but auto-switches the upstream to Zen for image requests, because mimo-v2.5-free is Zen-only. Pinning a model in the URL (e.g. /go/deepseek-v4-flash) keeps the request on Go even with images.

No configuration needed — it just works as long as you have an OpenCode Go or Zen subscription.

Default Text Model

Text-only requests (no images) are routed to DeepSeek V4 Flash (deepseek-v4-flash) by default — a fast, cheap chat model on both OpenCode Go and Zen. If deepseek-v4-flash returns a retriable failure (5xx, 429, or an empty/malformed body), the proxy automatically falls back to the vision model on Zen so you don't get stuck on transient errors. Pin a model in the URL (e.g. /go/kimi-k2.6) to opt out of the auto-default and the fallback.

Free Models

We support a pay-as-you-go model. Below are the prices per 1M tokens for completely free models available through OpenCode Zen.

Model	Model ID	Input	Output	Cached Read
Big Pickle	`big-pickle`	Free	Free	Free
DeepSeek V4 Flash Free	`deepseek-v4-flash-free`	Free	Free	Free
MiMo-V2.5 Free	`mimo-v2.5-free`	Free	Free	Free
Nemotron 3 Super Free	`nemotron-3-super-free`	Free	Free	Free

These models are available at https://opencode.ai/zen/v1/chat/completions via the /zen prefix. For the full model list and latest pricing, see the OpenCode Zen endpoint docs.

Set Up In Claude

The proxy now defaults text requests to deepseek-v4-flash and image requests to mimo-v2.5-free. The recommended base URL is /go: text runs on Go (https://opencode.ai/zen/go/v1), and image requests auto-switch to Zen so the vision model is reachable. If you want to pin a different model, use the URL-override trick described in Model Name Override.

Deploy this Worker to Cloudflare.
Copy your deployed Worker URL.
In Claude, open Configure third-party Inference.
Choose the gateway / third-party inference option.
Set the base URL to YOUR_DEPLOYED_WORKER_URL/go.
Set the auth scheme to x-api-key.
Paste your OpenCode API key.
Add deepseek-v4-flash as the model name (the proxy will route this to the text model — pin a different ID in the URL to override).

Important: do not add /v1/messages to the URL. Claude adds that path automatically.

Quick Claude Configuration

Use these values in Claude's Configure third-party Inference screen:

Setting	Value
Provider	Gateway / third-party inference gateway
Base URL	`YOUR_DEPLOYED_WORKER_URL/go`
Auth scheme	`x-api-key`
API key	Your OpenCode API key
Models	Add manually, for example `deepseek-v4-flash` (text) or `mimo-v2.5-free` (vision)

For the default example above, use /go so the default text model runs on Go. Image requests auto-switch to Zen under the hood. Use /zen instead if you want both text and image to stay on Zen. Do not add /v1/messages yourself. Claude adds the API path automatically.

To pin a different model without the smart-routing defaults, append the model ID after the prefix: YOUR_DEPLOYED_WORKER_URL/go/kimi-k2.6 uses kimi-k2.6 on Go directly and skips the default + auto-route + fallback. See Model Name Override for the full behavior matrix.

What This Does

The Worker accepts Claude's Anthropic-style requests at /v1/messages, converts them to OpenAI-style requests, and sends them to OpenCode Go by default.

You can choose an OpenCode upstream by adding a prefix to the Worker URL:

Worker URL suffix	Upstream
no suffix	OpenCode Go
`/go`	OpenCode Go
`/zen`	OpenCode Zen

For example, use YOUR_DEPLOYED_WORKER_URL/go for Go models and YOUR_DEPLOYED_WORKER_URL/zen for Zen models.

It also handles tool calls, streaming, and DeepSeek reasoning output so coding-agent workflows work correctly.

Important: this proxy has been live-tested with deepseek-v4-flash (default text) and mimo-v2.5-free (default vision) on the Zen path. Other OpenCode Go models are included from the public OpenCode Go model list, but provider behavior can vary, especially around streaming usage/token accounting.

Important Zen Limitation

OpenCode Zen support is partial.

This proxy currently works with Zen models that use the OpenAI-compatible /chat/completions endpoint.

Known Zen model categories that should work through /zen:

Zen model category	Examples
OpenAI-compatible chat models	`minimax-m2.7`, `minimax-m2.5`, `mimo-v2.5-free`, `glm-5.1`, `glm-5`, `kimi-k2.5`, `kimi-k2.6`, `grok-build-0.1`, `big-pickle`, `deepseek-v4-flash`, `deepseek-v4-flash-free`, `nemotron-3-super-free`

Known Zen model categories that do not work yet through this proxy:

Zen model category	Why it does not work yet
GPT models such as `gpt-5.5`, `gpt-5.5-pro`, `gpt-5.4`, `gpt-5.4-pro`, `gpt-5.3-codex`, `gpt-5.2`	Zen exposes these through `/responses`, and this proxy does not yet translate Anthropic Messages to OpenAI Responses API.
Claude models such as `claude-opus-4-8`, `claude-sonnet-4-6`, `claude-haiku-4-5`	Zen exposes these through `/messages`; this proxy's `/zen` Claude path currently translates to OpenAI-compatible `/chat/completions`.
Qwen models such as `qwen3.7-max`, `qwen3.6-plus`, `qwen3.5-plus`	Zen exposes these through `/messages` (Anthropic-compatible), not `/chat/completions`.
Gemini models such as `gemini-3.5-flash`, `gemini-3.1-pro`, `gemini-3-flash`	Zen exposes these through model-specific endpoints, not the generic chat-completions path used here.

Use /go for OpenCode Go. Use /zen only for Zen models listed as OpenAI-compatible chat models in the OpenCode Zen endpoint docs.

For Developers (OpenCode Cowork Proxy Worker)

Technically, this is a Cloudflare Worker gateway that lets Anthropic/Claude clients talk to OpenAI-compatible APIs, and lets OpenAI clients talk to Anthropic-compatible APIs.

The default upstream is OpenCode Go:

https://opencode.ai/zen/go/v1

This means Claude can be configured to use OpenCode Go models through this proxy without additional server-side configuration.

Claude / Anthropic SDK  ->  /v1/messages           ->  OpenAI-compatible upstream
OpenAI SDK              ->  /v1/chat/completions   ->  OpenAI-compatible upstream
OpenAI SDK              ->  /v1/chat/completions   ->  Anthropic upstream with x-upstream-format: anthropic

Detailed Claude Setup

Use Claude's Configure third-party Inference flow and add the Worker as a custom gateway.

Configure the gateway like this:

Setting	Value
Base URL	Your deployed Worker URL, or add `/go` or `/zen`
Auth scheme	`x-api-key`
API key	Your OpenCode Go API key
Model	Add manually, for example `deepseek-v4-pro`

Do not include /v1/messages in the Claude base URL. Claude will call /v1/messages; the Worker handles that path.

Use /go for OpenCode Go subscription models. Use /zen only for OpenCode Zen models available through the OpenAI-compatible /chat/completions endpoint. Zen GPT /responses, Zen Claude /messages, and Zen Gemini model-specific endpoints are not supported yet.

Manual Model Setup

Claude may not discover the OpenCode Go models automatically. Add the model manually in Configure third-party Inference.

Common OpenCode Go model IDs:

Model	Model ID	Upstream API style
GLM-5.1	`glm-5.1`	OpenAI-compatible
GLM-5	`glm-5`	OpenAI-compatible
Kimi K2.5	`kimi-k2.5`	OpenAI-compatible
Kimi K2.6	`kimi-k2.6`	OpenAI-compatible
DeepSeek V4 Pro	`deepseek-v4-pro`	OpenAI-compatible
DeepSeek V4 Flash	`deepseek-v4-flash`	OpenAI-compatible
MiMo-V2.5-Pro	`mimo-v2.5-pro`	OpenAI-compatible
MiMo-V2.5	`mimo-v2.5`	OpenAI-compatible
MiniMax M3	`minimax-m3`	Anthropic-compatible upstream
MiniMax M2.7	`minimax-m2.7`	Anthropic-compatible upstream
MiniMax M2.5	`minimax-m2.5`	Anthropic-compatible upstream
Qwen3.7 Max	`qwen3.7-max`	Anthropic-compatible upstream
Qwen3.6 Plus	`qwen3.6-plus`	Anthropic-compatible upstream

For the latest list, see the OpenCode Go endpoint docs:

https://opencode.ai/docs/go/#endpoints

For OpenCode's own config files, model IDs use the opencode-go/<model-id> format. For Claude's third-party inference setup through this proxy, use the raw API model ID such as deepseek-v4-pro, kimi-k2.6, or minimax-m3.

`claude.json` Example

You can also configure Claude with a claude.json gateway entry. Replace the Worker URL and API key with your own values.

{
  "inferenceProvider": "gateway",
  "inferenceGatewayBaseUrl": "YOUR_DEPLOYED_WORKER_URL/go",
  "inferenceGatewayApiKey": "YOUR_OPENCODE_GO_API_KEY",
  "inferenceGatewayAuthScheme": "x-api-key",
  "inferenceModels": [
    {
      "name": "glm-5.1"
    },
    {
      "name": "glm-5"
    },
    {
      "name": "kimi-k2.5"
    },
    {
      "name": "kimi-k2.6"
    },
    {
      "name": "deepseek-v4-pro"
    },
    {
      "name": "deepseek-v4-flash"
    },
    {
      "name": "mimo-v2.5-pro"
    },
    {
      "name": "mimo-v2.5"
    },
    {
      "name": "minimax-m3"
    },
    {
      "name": "minimax-m2.7"
    },
    {
      "name": "minimax-m2.5"
    },
    {
      "name": "qwen3.7-max"
    },
    {
      "name": "qwen3.6-plus"
    }
  ]
}

Deploy On Cloudflare

This project is intended to run as a Cloudflare Worker. Deploy it to Cloudflare using either the deploy button above or Cloudflare's Git-based Worker deployment flow.

Use these settings when connecting the repository in Cloudflare:

Setting	Value
Build command	empty
Deploy command	`npm run deploy`
Production branch	`main`

Do not deploy this as a normal Node.js web app. wrangler deploy builds and publishes the Worker from wrangler.toml.

Configuration

The Worker is zero-config by default. It forwards to OpenCode Go using OpenAI-compatible format. You can also route to OpenCode Zen by adding /zen to the Worker URL.

Optional request headers:

Header	Default	Description
`x-upstream-url`	`https://opencode.ai/zen/go/v1`	Upstream API base URL
`x-upstream-format`	`openai`	Upstream format: `openai` or `anthropic`
`x-api-key`	required	Upstream API key
`authorization`	optional	`Bearer <key>` also works
`anthropic-version`	`2023-06-01`	Forwarded when calling Anthropic-compatible upstreams
`anthropic-beta`	unset	Forwarded when calling Anthropic-compatible upstreams

The API key is validated locally before any upstream call. Missing or short keys receive a 401 response.

Prefix routes:

Path prefix	Upstream base URL
`/go`	`https://opencode.ai/zen/go/v1`
`/zen`	`https://opencode.ai/zen/v1`

Model Name Override

Claude Desktop may reject model names that don't look like Anthropic models (e.g. claude-sonnet-4-5 or anthropic/claude-*). To work around this, embed the real model name in the URL path after the prefix:

YOUR_DEPLOYED_WORKER_URL/zen/mimo-v2.5-free   # free Zen models
YOUR_DEPLOYED_WORKER_URL/go/deepseek-v4-pro   # paid Go models
YOUR_DEPLOYED_WORKER_URL/go/kimi-k2.6         # paid Go models, skip all smart routing

Claude appends /v1/messages, so the full request becomes YOUR_WORKER_URL/zen/mimo-v2.5-free/v1/messages. The proxy extracts the model from the path and uses it regardless of what Claude sends in the request body.

What URL pinning skips (use it when you want full control):

Behavior	Without URL pin	With URL pin (`/go/kimi-k2.6`)
Default text model	Forced to `deepseek-v4-flash`	Uses the pinned model
Image → Zen auto-route	`/go` image requests auto-switch to Zen upstream	Stays on `/go` even with images
Fallback on failure	Tries `mimo-v2.5-free` on Zen if `deepseek-v4-flash` fails	Returns the original error to the client

Usage:

Configure Claude with any Anthropic-looking model name (e.g. claude-sonnet-4-5-20250514) — this passes Claude's client-side validation.
Set the base URL to YOUR_WORKER_URL/zen/REAL_MODEL_ID (replace REAL_MODEL_ID with the actual OpenCode model).
The proxy silently maps the model for the upstream request.
The response uses the original model name you configured, so Claude sees consistency.

Setting	Value
Base URL	`YOUR_WORKER_URL/zen/mimo-v2.5-free`
Auth scheme	`x-api-key`
API key	Your OpenCode API key
Model	`claude-sonnet-4-5-20250514` (any Anthropic-looking name)

This works with all Go and Zen models.

API Endpoints

Path	Method	Purpose
`/v1/messages`	POST	Anthropic Messages API. Translates to OpenAI format by default.
`/v1/chat/completions`	POST	OpenAI Chat Completions API. Pass-through by default.
`/v1/models`	GET	Model discovery proxy.

OpenAI SDK Usage

Point any OpenAI-compatible client at the gateway. By default, /v1/chat/completions passes through to OpenCode Go.

from openai import OpenAI

client = OpenAI(
    base_url="YOUR_DEPLOYED_WORKER_URL/v1",
    api_key="your-opencode-go-api-key",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Hello"}],
)

OpenAI SDK To Anthropic

Set x-upstream-format: anthropic and point x-upstream-url at an Anthropic-compatible API.

curl YOUR_DEPLOYED_WORKER_URL/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_ANTHROPIC_KEY" \
  -H "x-upstream-url: https://api.anthropic.com" \
  -H "x-upstream-format: anthropic" \
  -d '{"model":"claude-sonnet-4-20250514","messages":[{"role":"user","content":"Hello"}]}'

Translation Notes

The gateway handles:

Anthropic Messages requests to OpenAI Chat Completions requests
OpenAI Chat Completions responses to Anthropic Messages responses
Tool calls and tool results in both directions
Streaming SSE in both directions
DeepSeek/OpenAI reasoning_content as Anthropic thinking blocks
Prompt cache key injection for OpenAI-style prefix caching

Prompt Caching

When translating Anthropic to OpenAI, the gateway injects a prompt_cache_key derived from a hash of the system prompt. This keeps requests with the same system prompt routed to the same backend node when the upstream supports OpenAI-style prefix caching.

Cache hit tokens from OpenAI-compatible usage metadata are mapped back to Anthropic's cache_read_input_tokens field.

Troubleshooting

`401 Invalid API key` (local gate)

{"error":{"type":"authentication_error","message":"Invalid API key: must be at least 32 characters."}}

Cause: The proxy validates the API key locally before any upstream call. It must be at least 32 characters and present in x-api-key or Authorization: Bearer <key>.

Fix:

Check the key is in the right header: x-api-key, not Authorization without Bearer.
Get a fresh key from https://opencode.ai/auth if the current one is shorter than 32 chars.
Test directly: curl -H "x-api-key: <KEY>" https://YOUR_WORKER/zen/v1/models.

`401 CreditsError` from Zen path

{"type":"error","error":{"type":"CreditsError","message":"Insufficient balance."}}

Cause: /zen and /go use different billing on OpenCode. Go credits don't transfer to Zen. The free models on Zen (e.g. mimo-v2.5-free) have a separate rate limit; all other Zen models need paid credits.

Fix:

Add Zen credits at https://opencode.ai/workspace → Billing.
For free usage only, stay on /go and pin models like kimi-k2.6 or deepseek-v4-flash (Go tier).
For image requests, the proxy auto-routes /go image requests to /zen for mimo-v2.5-free; if Zen is exhausted those will fail with 429 instead.

`429 FreeUsageLimitError` on free models

{"type":"error","error":{"type":"FreeUsageLimitError","message":"Rate limit exceeded."}}

Cause: OpenCode's free models (mimo-v2.5-free, deepseek-v4-flash-free, big-pickle, north-mini-code-free, nemotron-3-ultra-free) reset on a per-account basis. mimo-v2.5-free is the only vision-capable free model and is shared across all free users.

Fix:

Wait for the rate limit to reset (usually minutes to an hour).
Pin a paid model in the URL: https://YOUR_WORKER/go/kimi-k2.6 skips mimo-v2.5-free for vision.
Subscribe to Zen credits for guaranteed availability.

Empty response: only `thinking` block, no `text`

{
  "content": [{"type": "thinking", "thinking": "..."}],
  "stop_reason": "max_tokens"
}

Cause: Reasoning-capable models (deepseek-v4-flash, kimi-k2.6, minimax-m3) emit long reasoning_content blocks before the final answer. If max_tokens is too small, the model exhausts the budget during thinking and never reaches the actual text.

Fix:

Increase max_tokens in your request. Start with 500–1000 for reasoning models.
If you're using Claude Desktop, set the max output tokens in Settings → Model.
If you want a non-reasoning default, pin a model in the URL: /go/kimi-k2.6 is faster for short tasks.

Streaming hangs or first byte is slow

Cause: Cloudflare Workers can have a cold start on the first request after deploy (1-3 seconds). Subsequent requests are fast.

Fix:

Warm up the worker with a low-stakes request after deploy.
The proxy supports streaming for all /v1/messages requests; if streaming is broken on the client, check that stream: true is in the request body.

`count_tokens` returns wrong estimate

Cause: The proxy short-circuits /v1/messages/count_tokens with a local heuristic (chars/4 + 85 per image). It's a lower bound on purpose — Claude Desktop only uses this to display a context-window progress bar, not for billing.

Fix:

If you need exact token counts, call the upstream API directly.
The estimate is intentionally conservative so the progress bar doesn't jump past actual usage.

Image request fails on `/go`

Cause: The proxy auto-routes /go image requests to the /zen upstream (since mimo-v2.5-free is Zen-only). If Zen has no credits or the free model is rate-limited, the image will fail.

Fix:

Use /zen URL directly if you have Zen credits.
Pin a vision-capable model in the URL: /go/qwen3.6-plus stays on Go and uses a vision model (note: requires Go tier support for qwen3.6-plus — check the Go model list).
Wait for the free vision model rate limit to reset.

`404` or model not found after URL pin

Cause: The model ID in the URL path is unknown to the upstream, or it's only available on one of /go / /zen.

Fix:

Check the model exists in the relevant list:
- Go models
- Zen models
Some models (e.g. gpt-5.5, claude-opus-4-8) use non-/chat/completions endpoints and aren't supported by this proxy yet.

Request times out after 30s

Cause: Cloudflare Workers have a 30-second CPU time limit on the free plan and a 30-second wall time on paid plans for streaming responses.

Fix:

Reduce max_tokens for the task.
Switch to a faster, non-reasoning model: /go/kimi-k2.6 or /zen/glm-5.
For long tasks, the worker can be deployed on a paid plan with extended limits.

`503` Worker exceeded resource limits (Error 1102)

Error 1102: Worker exceeded resource limits (CPU time or memory) and was terminated.

Cause: Cloudflare Workers on the Free plan have a 10ms CPU time limit per request. Long reasoning streams (e.g. deepseek-v4-flash, kimi-k2.6) emit thousands of small SSE deltas per second; each enqueue + JSON.stringify costs CPU and adds up fast. The proxy now batches consecutive same-type deltas into 1KB chunks to stay under the limit, but extreme requests (e.g. very long thinking with 10K+ tokens) can still exceed it.

Fix:

This proxy already batches deltas (see src/translate/stream/openai-to-anthropic.ts and anthropic-to-openai.ts) — make sure you're running a deployed version newer than commit e4303d1 / a36ec80.
Reduce max_tokens for very long reasoning tasks. Reasoning models use budget on the thinking block first; with low max_tokens the model exhausts it during thinking and never reaches the answer (see "Empty response: only thinking block" above).
Switch to a non-reasoning model for short tasks: /go/kimi-k2.6 (Go tier) or /zen/glm-5 (Zen tier).
For persistent 1102 errors on Free plan, upgrade to Workers Paid ($5/mo) — raises the CPU limit to 30s.
The same request always hits the same limit, so retrying won't help.

Development

npm install
npm test               # unit tests (no API key required)
npm run e2e            # E2E smoke test against deployed worker (requires OPENCODE_API_KEY)
npm run deploy -- --dry-run

Project structure:

src/
├── index.ts                          Main Worker router and auth gate
├── auth.ts                           API key extraction and validation
├── cache.ts                          Prompt cache key utilities
└── translate/
    ├── request/                      Request translators
    ├── response/                     Response translators
    └── stream/                       SSE stream translators
test/
├── auth.test.ts
├── cache.test.ts
├── index.test.ts
├── request.test.ts
├── response.test.ts
└── stream.test.ts

License

MIT. See LICENSE.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
scripts		scripts
src		src
test		test
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json
vitest.config.ts		vitest.config.ts
wrangler.toml		wrangler.toml

Navigation Menu

Repository files navigation

OpenCode Cowork Proxy Worker

Image / Vision Support

Default Text Model

Free Models

Set Up In Claude

Quick Claude Configuration

What This Does

Important Zen Limitation

For Developers (OpenCode Cowork Proxy Worker)

Detailed Claude Setup

Manual Model Setup

`claude.json` Example

Deploy On Cloudflare

Configuration

Model Name Override

API Endpoints

OpenAI SDK Usage

OpenAI SDK To Anthropic

Translation Notes

Prompt Caching

Troubleshooting

`401 Invalid API key` (local gate)

`401 CreditsError` from Zen path

`429 FreeUsageLimitError` on free models

Empty response: only `thinking` block, no `text`

Streaming hangs or first byte is slow

`count_tokens` returns wrong estimate

Image request fails on `/go`

`404` or model not found after URL pin

Request times out after 30s

`503` Worker exceeded resource limits (Error 1102)

Development

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

OpenCode Cowork Proxy Worker

Image / Vision Support

Default Text Model

Free Models

Set Up In Claude

Quick Claude Configuration

What This Does

Important Zen Limitation

For Developers (OpenCode Cowork Proxy Worker)

Detailed Claude Setup

Manual Model Setup

claude.json Example

Deploy On Cloudflare

Configuration

Model Name Override

API Endpoints

OpenAI SDK Usage

OpenAI SDK To Anthropic

Translation Notes

Prompt Caching

Troubleshooting

401 Invalid API key (local gate)

401 CreditsError from Zen path

429 FreeUsageLimitError on free models

Empty response: only thinking block, no text

Streaming hangs or first byte is slow

count_tokens returns wrong estimate

Image request fails on /go

404 or model not found after URL pin

Request times out after 30s

503 Worker exceeded resource limits (Error 1102)

Development

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

`claude.json` Example

`401 Invalid API key` (local gate)

`401 CreditsError` from Zen path

`429 FreeUsageLimitError` on free models

Empty response: only `thinking` block, no `text`

`count_tokens` returns wrong estimate

Image request fails on `/go`

`404` or model not found after URL pin

`503` Worker exceeded resource limits (Error 1102)