droid-oauth-opencode

Use Factory (Droid) proxied models inside OpenCode, authenticated with the credential the Factory droid CLI already stores on your machine. No API key to paste, no local proxy server, no opencode auth login step.

Two providers are registered:

| OpenCode provider | Route | Models | | --- | --- | --- | | factory | /api/llm/a/v1 (Anthropic Messages API; SDK appends /messages) | Claude Opus 4.8/4.7/4.6/4.5 (+ Fast Modes), Fable 5, Sonnet 5/4.6/4.5, Haiku 4.5 | | factory-g | /api/llm/o/v1 (OpenAI chat/completions) | GLM-5.2 Fast, GLM-5.2, Kimi K2.7 Code, Kimi K2.6, MiniMax M3, MiniMax M2.7, DeepSeek V4 Pro |

factory-g/glm-5.2-fast is Factory's "Droid Core" GLM-5.2 running on Fireworks — the fast variant that has no equivalent in the stock Z.AI / Devin providers. Anthropic Opus *-fast variants send speed: "fast" and the fast-mode-2026-02-01 anthropic-beta flag (same as the droid CLI).

Install

npx droid-oauth-opencode setup --global

Then restart OpenCode and pick a factory/* or factory-g/* model. The plugin reuses the droid CLI credential automatically. If you have never logged in with the droid CLI, run droid once first (or export FACTORY_API_KEY=fk-...).

Project-local install (writes ./opencode.json instead of the global config):

npx droid-oauth-opencode setup --project

How it works

Both routes live on https://app.factory.ai behind one WorkOS bearer token. The plugin registers one provider per route and attaches a custom fetch that injects auth + Factory routing headers.

Authentication and token refresh

The droid CLI stores its WorkOS OAuth tokens encrypted at ~/.factory/auth.v2.file (AES-256-GCM, envelope iv:tag:ciphertext in base64, key base64 in ~/.factory/auth.v2.key) as JSON { access_token, refresh_token, active_organization_id }.

Access tokens are short-lived JWTs. On each request the plugin:

1. Decrypts the credential and reads the access-token expiry.
2. If expired (or within 60s of expiry), exchanges the refresh token at https://api.workos.com/user_management/authenticate (grant_type=refresh_token, Factory's production client_id, no organization_id — WorkOS returns organization_not_found if you send one), then re-encrypts the rotated credential back into the exact same envelope (atomic write, so a concurrent droid read never sees a half-written file).
3. If the server still returns 401, forces one more refresh and retries the request once.

Set FACTORY_API_KEY=fk-... to bypass all of this with a long-lived key.

The factory-g (OpenAI) route quirks

The generic route is fussier than the Anthropic one. sanitizeBody / factoryGenericFetch normalize four things discovered by probing the live endpoint:

1. x-api-provider header must be the model's upstream backend (fireworks), not the registry enum generic-chat-completion-api (which 400s with Invalid x-api-provider header).
2. User-Agent must be factory-cli/0.57.0 or newer; older versions 400 with too old to use this endpoint.
3. Body leakage — the @ai-sdk/openai-compatible SDK leaks provider options (baseURL, apiKey) into the JSON request body, which the route rejects. Unknown keys are stripped to a chat/completions allow-list.
4. role: "system" is rejected by the Fireworks route; messages are rewritten to role: "developer".
5. WAF phrases — Factory's edge 403s on exact OpenCode identity strings. Softened in-proxy (punctuation only, same meaning):
   • You are powered by the model named → You are powered by the model, named
   • You are OpenCode, the best coding agent on the planet. → You are OpenCode. The best coding agent on the planet.
6. Skills beta — OpenCode often sends skills-* without a code_execution tool; Anthropic returns 400. The proxy drops that beta unless the tool is present.

Endpoints

• Anthropic: POST https://app.factory.ai/api/llm/a/v1/messages
• Generic: POST https://app.factory.ai/api/llm/o/v1/chat/completions
• Token refresh: POST https://api.workos.com/user_management/authenticate

Environment variables

| Variable | Purpose | | --- | --- | | FACTORY_API_KEY | Long-lived fk-... key; bypasses the CLI credential + refresh entirely. | | FACTORY_HOME | Override the ~/.factory directory. | | FACTORY_API_BASE_URL | Override https://app.factory.ai. | | FACTORY_CLI_VERSION | Override the factory-cli/<ver> User-Agent (default 0.57.0). | | FACTORY_WORKOS_BASE_URL | Override the WorkOS user_management base URL. | | FACTORY_WORKOS_CLIENT_ID | Override the WorkOS client id used for refresh. |

Development

bun install
bun run build     # tsc -> dist/, bundles bin/setup.js
bun test          # smoke test: decrypt credential + one live request

Source layout:

• src/credentials.ts — decrypt / refresh / re-encrypt the Factory credential.
• src/models.ts — static model catalogs for both providers.
• src/index.ts — the plugin: provider registration + the two fetch hooks.
• bin/setup.ts — the npx installer that patches OpenCode's config.

The model catalogs and route quirks were reverse-engineered from the droid binary's embedded model registry and confirmed against the live API.

Notes

• Model availability, -fast variants, and pricing are gated by Factory feature flags on your account; the picker only shows what your org can use.
• HTTP 402 ("reached your 5-hour standard usage limit") is a Factory billing limit, not a plugin bug — the request authenticated fine.
• This project is unaffiliated with Factory; it just reuses the credential the official droid CLI already stores locally.

License

MIT — see LICENSE.