@magpie-community/opencode-factory-auth

Signs in to a Factory (Droid) subscription and sends model requests to Factory's API the way droid sends them. Provider id: factory.

Sign-in

• Sign in with Factory (device code). This is WorkOS's device flow under droid's own client. The browser opens Factory's page with the code filled in. Confirm it there, and the plugin picks up the tokens.
• Factory API key. A key (fk-…), as droid takes one from FACTORY_API_KEY. It is sent as the bearer token with droid's headers and never renewed. A key carries no organization header; whoami, asked once with the key, says whose it is and where its organization is served.
• Organization. A token that isn't in an organization yet is put in the first one your account belongs to.
• Whoami. The plugin asks whoami for your active organization, its region (EU orgs go to api.eu.factory.ai) and any host of the org's own. Every request then carries them, as droid's do.

Where the sign-in is kept

The sign-in is kept wherever the host keeps provider sign-ins:

• OpenCode: ~/.local/share/opencode/auth.json
• magpie: plugin-auth.json

It holds WorkOS's access and refresh tokens, with the organization, region and host.

Refreshing.

• magpie renews the access token three minutes before it lapses, through the plugin's auth.refresh, once for the account and before its requests and usage need it; magpie saves the new tokens. OpenCode doesn't call that hook: there the token is renewed two minutes before it lapses, before a request, and the plugin saves it.
• WorkOS rotates the refresh token, so the plugin never runs two refreshes at once, and never spends one refresh token twice: a request that comes before magpie has saved a renewal goes on with the new tokens.
• The account counts as signed out only when WorkOS refuses the refresh token (a 4xx other than 429). Any other failure keeps the sign-in, and the current token is used while it lasts.

Refusals. If Factory refuses the organization a request names, the plugin asks whoami again and resends the request once. If the refusal stands, the error says what to check.

Requests from other agents

Factory serves a subscription's model requests to Droid. Every request droid sends opens its system prompt with "You are Droid, an AI software engineering agent built by Factory.", so another agent's request to Factory's OpenAI-shaped API (/api/llm/o: GPT and Grok on Responses, the open models on chat completions) opens with that line too, the agent's own prompt after it. On Anthropic's Messages (/api/llm/a), the plugin adapts fixed client metadata while preserving the task instructions and history; see Claude Code through magpie below.

Models

Each model is served on the one API droid uses for it:

| API | Models | |---|---| | Anthropic Messages (/api/llm/a) | Fable 5.1, Fable 5, Opus 5.5, Opus 5, Opus 4.8 (each also in Fast Mode), Sonnet 5.5, Sonnet 5, Sonnet 4.6, Haiku 5.5, Haiku 4.5, MiniMax M2.7 | | OpenAI Responses (/api/llm/o/v1) | GPT-6.1 Sol, GPT-6 Sol/Astra/Luna, GPT-5.6 Sol/Terra/Luna, GPT-5.5, GPT-5.4, GPT-5.3-Codex (each also in Fast Mode), Grok 4.7, Grok 4.6 | | Chat completions (/api/llm/o/v1) | GLM-5.3, GLM-5.3-Flash, GLM-5.2, Kimi K3, DeepSeek V4.1 Flash, Qwen3.8 Max, MiniMax M3, Mistral Large 4, Mistral Medium 3.5, Nemotron 3 Ultra | | Gemini generate (/api/llm/g/v1/generate) | Gemini 4 Argon, Gemini 3.1 Pro, Gemini 3.8/3.7/3.6/3.5 Flash, Gemini 3 Flash |

Reasoning efforts are the variants droid offers for each model. DeepSeek V4.1 Flash, Kimi K3 and GLM-5.2 have none, droid's "off": they stop thinking. Reasoning is mandatory on GLM-5.3 and GLM-5.3-Flash (Fireworks turns none away), so low is their least.

Gemini's requests go to Factory's generate route with the model in the body, droid's line opening the system instruction. Gemini 4 Argon is behind a Factory feature flag in droid 0.241.1, so an account it isn't on for may be refused it.

A Fast Mode model (claude-opus-5-5-fast, gpt-5.5-fast, …) is its base model served faster, for more of the plan's credits. It goes out as droid sends it: its own id, with speed: "fast" and the fast-mode-2026-02-01 beta on Anthropic's API, at service_tier: "priority" on OpenAI's. A request that asks the base model for fast (Claude Code's speed: "fast", a Responses request's service_tier: "priority") is sent to its Fast Mode model, as droid's own fast toggle does: asked of the base model, Factory answered at standard speed (#119). Opus 4.6/4.7 Fast, GPT-5.4 Mini Fast, GPT-5.6 Sol Fast (unreleased) and GLM-5.2 Fast (behind a feature flag) are not included.

auto is not included: droid picks it on the client side.

Requests too large

Factory refuses a request body over its size limit with 413 "Request Entity Too Large". The limit is on bytes, not on the model's context, and images reach it first. On a 413 the plugin sends the request once more with every image and file but the newest replaced by a line saying it was left out, each tool result kept beside its call, as droid does. If that is still too large, or there was nothing to leave out, the error says it is Factory's byte limit: compact the conversation or start a new one.

Use

magpie plugin add @magpie-community/opencode-factory-auth
magpie plugin login factory

Claude Code through magpie

Factory's Anthropic route requires Droid's fixed client preamble and refuses some of Claude Code's fixed environment, model and built-in skill metadata. The plugin adapts that metadata while preserving the coding instructions, environment values, tool definitions, images and reasoning options. Ordinary Droid requests remain unchanged, and OpenAI routes keep their existing request adapter. Quoted fixed metadata in tool results is handled as described below. The fixed Claude identity is recognized as a complete first line too when magpie joins it with the following system instructions into one text block.

For the built-in update-config skill, the adapter changes only the known self-reference "not Claude" to "not the assistant" inside a complete generated skill-list reminder. Its configuration instructions, other skill descriptions, tools and caller text are preserved.

Claude 5 can send runtime metadata in string-content system messages instead of user-message reminders. The plugin adapts the known environment, model and configuration-skill metadata in that shape too, preserving the message role, environment values and session instructions. Startup hooks (SessionStart or SubagentStart, including additional context) and deferred-tool announcements may precede that environment block. The adapter preserves the prefix verbatim and adapts the generated runtime context after it, including when folded into user text. After /model switches, Claude Code can send model-only system updates without the initial environment paragraph. Updates beginning with the known model paragraph and containing complete generated token metadata are adapted too, including accumulated switches when resuming a conversation. Token budgets, permission instructions and conversation history are preserved. The same model-update adaptation covers updates that start with the complete generated # Environment update block after a working-directory change.

Claude Code's generated compaction summaries also have a fixed opening that Factory refuses. The adapter recognizes the two generated opening sentences and rephrases only the first in user strings and text blocks, including summaries without a Summary: label. The generated artifact provenance marker and note stay verbatim. If the summary itself quotes any of the refused client phrases handled in tool output, the context after the two opening sentences is encoded as a JSON string with explicit decoding instructions. Decoding restores the exact summary, transcript path and continuation instructions; context without those phrases stays verbatim. Complete generated reminders for restored Read calls and omitted files are adapted too. Read arguments and file paths are preserved; quoted or incomplete reminders are left alone. Claude Code (2.1.291 and later) also restores skills invoked before compaction as one generated reminder. Factory refuses that fixed opening, and the restored skill text can quote the same client phrases as tool output. A complete reminder has its opening sentence rephrased. When its skill text quotes a refused phrase, that text is encoded as a JSON string with explicit skill-text decoding instructions; decoding restores the skill text exactly. Skill text without those phrases stays verbatim inside the rephrased reminder. Pasted fragments and incomplete reminders are left untouched. Claude Code can also combine reminders into a token-prefixed system turn without the reminder wrappers. The adapter handles the known metadata paragraphs in that form, including skill-list updates and complete invoked-skills reminders. A complete reminder is the fixed opening sentence followed by the complete IMPORTANT: paragraph; skill text then runs until the next known attachment (a token marker, hook, restored Read, omitted-file note, Read result, environment, model line, skill-list header, changed-file note, agent-type list or MCP server instructions) or the end of the turn. Numbered file contents, token markers and SessionStart hook output stay. The same adaptation applies when a translating gateway folds that complete token-prefixed block into a user message. Standalone skill updates can instead start with the generated skill-list header and end with a token marker, for example when leaving auto mode. That complete form receives the same adaptation while retaining the skill descriptions and mode-change instructions. The skill list can also come alone as a system message, with only its entries: no header, no reminder wrapper and no token marker, as text blocks or a string. A system message that opens with an entry gets the same change to the known update-config line; everything else in it stays. When runtime notifications precede the skill update, the adapter recognizes the complete skill-header and bullet-list paragraphs inside a token-terminated bundle. It preserves MCP connection errors, other notifications, mode updates and custom skill descriptions while adapting the known built-in description. After /compact, Claude Code also sends the files it restores and the runtime context as one turn that opens with an unwrapped restored-file note, Read call or complete invoked-skills reminder, with no token marker first. A complete skills reminder is recognized even when no files were restored and no environment or token paragraph follows it. A restored-file opening requires the generated environment paragraph, a token marker or a complete skills reminder before any hook output. Each fixed paragraph (file notes, Read calls and results, skills, environment, model line) is adapted as in a token-prefixed bundle; everything else stays byte-for-byte. Quoted, incomplete and hook-owned forms are left alone.

Factory can also refuse fixed client phrases quoted in tool results, such as the identity/environment definitions printed when inspecting this plugin's source, or the compaction opening embedded in a historical session JSONL record. Those text results are represented as JSON strings with explicit decoding instructions and Unicode escapes for the fixed phrases. Decoding the string recovers the exact original output, including quotes, backslashes and Unicode; the plugin does not delete the output or replace its identities with different ones. Tool ids, cache/error markers and non-text blocks stay unchanged. Ordinary tool results are forwarded as before. The same lossless encoding applies to restored or @-attached Read file text after the exact Result of calling the Read tool: header, including complete reminder wrappers and announced runtime bundles. The generated header and wrapper stay intact. File text without a known refused phrase is unchanged. Claude Code's complete changed-on-disk notifications receive the same handling: only the numbered file snippet is encoded, retaining its TAB or colon line separators, hunk separators and truncation notice. The path and instructions about respecting the current file stay verbatim. This covers wrapped user attachments and generated system bundles, including notifications before a token marker. Changed-file text supplied by hooks, quoted or incomplete notifications, and notices that omit the snippet are left alone. Model and environment metadata after non-startup hooks retain their existing adaptation.

Use magpie's Claude Code integration to select the Factory provider. magpie manages the provider-specific client settings, including capability and permission configuration. Fields such as safeguards and context_management are forwarded for Factory to validate; the request adapter preserves them along with the other request options. One exception: Factory's streaming route refuses context_management ("Extra inputs are not permitted") while its non-streaming route takes it, so a streamed request is sent without it. The server then clears nothing from the context; Claude Code's own compaction still runs.

The same metadata adaptation applies to /messages and /messages/count_tokens, so token counting sees the prompt used for inference. Serving the counting endpoint still depends on the host and upstream. Model and feature availability depend on the Factory account, organization region and upstream API. Connectivity has been verified with Sonnet 4.6, Sonnet 5.5 and Opus 5.5. Two-turn Read tool calls with Claude Code's default tool set have been verified with Sonnet 4.6 and Sonnet 5.5, including Sonnet 5.5 selected through a routing group.

Tool schema compatibility

Factory's Anthropic route rejects tool schemas with root-level anyOf, oneOf or allOf. The plugin nests those schemas under an internal arguments property instead of removing their branch constraints. It wraps examples and previous tool inputs too, then unwraps tool replies before returning them to OpenCode or magpie. Ordinary object schemas and OpenAI routes keep their existing behavior.

Local schema pointers are rebased to the nested schema; $id resources and anchors retain their resolution scope, and literal data is unchanged. Recursive schema resources without $id fail explicitly because nesting would change their scope. Affected streaming tool arguments are buffered until their content block is complete. Other events, tool ids, usage and upstream errors pass through; an incomplete or invalid argument envelope fails instead of being executed.