factory-droid-openai 1.3.3

Unofficial OpenAI-compatible HTTP bridge for Factory Droid

Factory Droid OpenAI Bridge

Use your Factory Droid models in VS Code, Cursor, Continue, Aider, Zed - any tool that speaks the OpenAI Chat Completions API.

This bridge runs locally on your machine. It translates between OpenAI's API format and Factory Droid, so your tools talk to http://127.0.0.1:8787/v1 like it's OpenAI. Models like GPT-5.4 and Gemini 3.1 Pro become available anywhere you can set a custom base_url. The bridge never runs tools and never reads your credentials.

Not affiliated with, endorsed by, or maintained by Factory.

Contents

• Quick start
• Client guides
  • VS Code (Copilot BYOK)
  • OpenClaw
  • Hermes Agent
  • OpenAI Python client
• API compatibility
• API reference
• Installation
  • Docker
• Configuration
• Features
• How it works
• Requirements
• Limits and metrics
• Warm sessions
• Logging
• Authentication
• Tool execution safety
• Bridge architecture limits
• Troubleshooting
• Development
• License

Quick start

Requires Python 3.11+, an installed droid CLI, and an authenticated Factory session. Verify the CLI once with droid --version.

Install the bridge as a tool, then run it from the project directory Droid should operate in. The working directory defaults to the current directory, so no environment variable is required for local use.

uv tool install factory-droid-openai
cd /path/to/your/project
factory-droid-openai

pipx install factory-droid-openai works the same way if you prefer pipx.

No droid CLI on the host? Use Docker - it bundles Droid inside the container, you only need FACTORY_API_KEY.

docker run -d --name droid-bridge \
  -p 127.0.0.1:8787:8787 \
  -e FACTORY_API_KEY="$FACTORY_API_KEY" \
  ghcr.io/mrwogu/factory-droid-openai:latest

From another terminal, send a request:

curl --fail http://127.0.0.1:8787/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gpt-5.4",
    "messages": [
      {"role": "user", "content": "Reply with hello."}
    ]
  }'

Use any model ID from GET /v1/models, or the factory-droid alias to pick Droid's configured default.

API compatibility

✅ Supported · ⚠️ Partial · ❌ Unsupported

OpenAI capability	Status	Bridge behavior
Text chat completions	✅	Returns one or more assistant choices
System and developer messages	✅	Serialized with the complete transcript
Non-streaming responses	✅	OpenAI-compatible JSON completion
Streaming responses	✅	SSE chunks followed by [DONE]
Function tool schemas	✅	Serialized into the strict Droid prompt
Tool choice	✅	auto, none, required, or one named function
Tool-result continuation	✅	Client resends the complete transcript on the next request
Parallel tool calls	✅	Several sequential tool calls per turn, bounded by a configured cap
Stop sequences	✅	stop truncates the reply and interrupts the Droid turn
Reasoning output	✅	Emitted as reasoning and reasoning_content
Token usage	✅	Includes cache read and write token details
Model selection	✅	/v1/models discovers the authenticated account's current catalog
Multimodal content	⚠️	Inline image and file data URIs use SDK attachments; remote URLs are rejected
Structured outputs	✅	json_schema and json_object are enforced by Droid and validated by the bridge
Sampling controls	❌	temperature, top_p, penalties, and seed are ignored
Multiple choices	✅	n runs that many Droid turns sequentially and returns n choices
Log probabilities	❌	logprobs and top_logprobs are ignored
Output token limits	❌	max_tokens and max_completion_tokens are ignored
Stored completions	❌	store, metadata, listing, retrieval, and deletion are unavailable
Prompt cache controls	❌	OpenAI cache keys and retention settings are ignored
Built-in web search	❌	web_search_options is ignored
Audio output	❌	Audio modalities and audio response fields are ignored

Unsupported API families

API family	Status
Responses API	❌
Embeddings	❌
Images	❌
Audio	❌
Files and batches	❌
Fine-tuning	❌
Moderations	❌
Realtime	❌
Vector stores and uploads	❌

Features

• OpenAI-compatible POST /v1/chat/completions
• Dynamic OpenAI-compatible GET /v1/models
• Non-streaming JSON responses
• Streaming server-sent events with [DONE] termination
• System, developer, user, assistant, and tool message mapping
• Function tool schemas and validated tool-call responses
• Reasoning delta and token usage mapping
• Optional constant-time bearer authentication
• Bounded request timeout, payload size, and process concurrency
• Queue-based admission control and Prometheus-style metrics
• Warm Droid session pool and detached teardown for low request latency
• Structured phase-timed logging in text or JSON
• Inline image and document attachments over the native SDK channel
• Stop sequences, n choices, and multiple tool calls per turn
• Optional Droid session continuity across requests
• Guarded context, compaction, fork, rename, and close session extensions
• Client disconnect cancellation
• Discoverable, verified Factory-native tool disabling
• Versioned and externally validated OpenAPI 3.1 contract
• Strict typing, locked dependencies, and at least 95% branch coverage

How it works

OpenAI client
    |
    | POST /v1/chat/completions
    v
Factory Droid OpenAI Bridge
    |
    | OpenAI transcript + tool schemas encoded as JSON
    v
droid-sdk-python
    |
    | JSON-RPC over a local droid exec subprocess
    v
Factory Droid

Factory's SDK controls a complete Droid agent rather than exposing a raw model inference endpoint. The bridge uses a fresh Droid session for each completion, serializes the complete OpenAI transcript into one prompt, and maps Droid stream events back to OpenAI response objects. Sessions are started ahead of time and kept warm, so the startup cost stays off the request path; see Warm sessions.

External tools use a strict text protocol. Droid emits:

<tool_call>{"name":"weather","arguments":{"city":"Gdansk"}}</tool_call>

The bridge validates the tool name and JSON arguments, then returns a standard OpenAI tool_calls object. The OpenAI client executes the tool and sends its result in the next request.

Requirements

• Python 3.11 or newer
• Installed Factory droid CLI
• Authenticated Droid session or FACTORY_API_KEY

Verify the CLI before starting the bridge:

droid --version

Run Droid normally once if authentication or first-time setup is required. For service-account execution, export FACTORY_API_KEY; the Droid subprocess inherits it directly. The bridge does not rename or copy that credential.

Python

Python 3.11 or newer (3.12, 3.13, and 3.14 are also tested). On Windows, the official installer works; ensure Python is on PATH or invoke it via py -3.

Factory Droid CLI

Install the Factory droid CLI:

OS	Command
macOS / Linux	curl -fsSL https://app.factory.ai/cli | sh
Windows (PowerShell)	irm https://app.factory.ai/cli/windows | iex
Any (npm)	npm install -g droid

Authenticate interactively once, or set a FACTORY_API_KEY service-account key generated at app.factory.ai/settings/api-keys.

Docker (optional)

Docker or Docker Desktop is optional and only needed to run the bridge as a long-lived container from the prebuilt image on GHCR. See the Docker section. The bridge itself has no Docker runtime dependency.

Installation

Released versions are published to PyPI.

uv

uv tool install factory-droid-openai
factory-droid-openai

pipx

pipx install factory-droid-openai
factory-droid-openai

pip

python3 -m venv .venv
.venv/bin/python -m pip install factory-droid-openai
.venv/bin/factory-droid-openai

From source

git clone https://github.com/mrwogu/factory-droid-openai.git
cd factory-droid-openai
uv sync
uv run factory-droid-openai

Docker

A prebuilt image is published to GitHub Container Registry for each release:

docker pull ghcr.io/mrwogu/factory-droid-openai:latest

The image bundles a pinned Droid CLI with auto-updates disabled, so the bridge and CLI versions are reproducible. Droid runs as a non-root user (uid 1000) and the bridge listens on 0.0.0.0:8787 inside the container.

Run it with a Factory service-account key and a bridge bearer token. Bind the port to loopback on the host so the bridge is not exposed to the network:

docker run -d --name droid-bridge \
  -p 127.0.0.1:8787:8787 \
  -e FACTORY_API_KEY="$FACTORY_API_KEY" \
  -e FACTORY_DROID_OPENAI_API_KEY="$FACTORY_DROID_OPENAI_API_KEY" \
  -v "$PWD":/work:ro \
  ghcr.io/mrwogu/factory-droid-openai:latest

docker compose up -d with the included docker-compose.yml does the same. Copy the compose file, set FACTORY_API_KEY and FACTORY_DROID_OPENAI_API_KEY in your environment or a .env file, and run:

export FACTORY_API_KEY=fk-...
export FACTORY_DROID_OPENAI_API_KEY="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
docker compose up -d

Podman works as a drop-in replacement. Use podman-compose or replace docker with podman in the commands above. The compose file uses ${VAR:?error} syntax to require environment variables; set both keys in your shell or in a .env file before starting the container.

The /work volume is optional. The bridge disables all Factory-native tools, so Droid never writes to the working directory; a read-only mount is safe and sufficient when you want Droid to see project context. Remove the -v flag when using docker run, or comment out the volumes block in docker-compose.yml when using Compose, to run without a workdir.

To build the image locally instead of pulling it:

docker build -t factory-droid-openai .
# or, pinning the Droid CLI version:
docker build --build-arg DROID_VERSION=0.174.0 -t factory-droid-openai .

Authentication inside a container uses FACTORY_API_KEY only. Interactive droid login does not work in a container, so generate a key at app.factory.ai/settings/api-keys and pass it as an environment variable. The bridge never reads or copies that credential; the Droid subprocess inherits it directly.

Never publish the bridge port to a non-loopback interface without a configured FACTORY_DROID_OPENAI_API_KEY. See SECURITY.md before any remote deployment.

The default address is http://127.0.0.1:8787.

Check health and inspect the configured alias:

curl --fail http://127.0.0.1:8787/health
curl --fail http://127.0.0.1:8787/v1/models

Stream a completion:

curl --no-buffer --fail http://127.0.0.1:8787/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "gemini-3.1-pro-preview",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Count from one to three."}
    ]
  }'

OpenAI Python client

Install the official client in your application:

python -m pip install "openai>=2,<3"

Non-streaming

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8787/v1",
    api_key="none",
)

response = client.chat.completions.create(
    model="gpt-5.4",
    messages=[
        {"role": "system", "content": "Answer concisely."},
        {"role": "user", "content": "Explain JSON-RPC in one sentence."},
    ],
)

print(response.choices[0].message.content)

Structured output

Use OpenAI response_format with either json_schema or json_object:

response = client.chat.completions.create(
    model="gpt-5.4",
    messages=[{"role": "user", "content": "Pick an integer from 1 to 10."}],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "answer",
            "strict": True,
            "schema": {
                "type": "object",
                "properties": {"answer": {"type": "integer"}},
                "required": ["answer"],
                "additionalProperties": False,
            },
        },
    },
)

The bridge sends the schema through Droid's native outputFormat RPC field, then independently parses and validates the completed JSON. Invalid JSON, duplicate keys, non-finite numbers, schema violations, and trailing output fail closed. Remote JSON Schema references are rejected. response_format cannot be combined with bridge text tools or stop. Structured streaming buffers the JSON until it validates, then emits it in one content chunk, so no content chunks arrive while the model is generating. Buffered output is capped by FACTORY_DROID_OPENAI_MAX_STRUCTURED_OUTPUT_BYTES, and a schema larger than FACTORY_DROID_OPENAI_MAX_TOOL_SCHEMA_BYTES is rejected with 413.

Streaming

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8787/v1",
    api_key="none",
)

stream = client.chat.completions.create(
    model="gemini-3.1-pro-preview",
    messages=[{"role": "user", "content": "Count from one to three."}],
    stream=True,
)

for chunk in stream:
    delta = chunk.choices[0].delta
    if delta.content:
        print(delta.content, end="", flush=True)

Function tools

import json

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8787/v1",
    api_key="none",
)

messages = [{"role": "user", "content": "What is the weather in Gdansk?"}]
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Read current weather for a city.",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"},
                },
                "required": ["city"],
                "additionalProperties": False,
            },
        },
    }
]

first = client.chat.completions.create(
    model="gpt-5.4",
    messages=messages,
    tools=tools,
)
assistant = first.choices[0].message
messages.append(assistant.model_dump(exclude_none=True))

for tool_call in assistant.tool_calls or []:
    arguments = json.loads(tool_call.function.arguments)
    result = {"city": arguments["city"], "temperature_c": 20}
    messages.append(
        {
            "role": "tool",
            "tool_call_id": tool_call.id,
            "content": json.dumps(result),
        }
    )

final = client.chat.completions.create(
    model="gpt-5.4",
    messages=messages,
    tools=tools,
)
print(final.choices[0].message.content)

Hermes Agent

Configure Hermes as a custom OpenAI-compatible provider in ~/.hermes/config.yaml:

model:
  provider: custom
  default: gpt-5.4
  base_url: http://127.0.0.1:8787/v1
  api_key: none
  api_mode: chat_completions

model.default is forwarded to Droid as an explicit model ID. Replace it with another ID from droid exec --help when needed.

Hermes retains ownership of tool execution. The bridge converts Droid's strict tool protocol into OpenAI tool_calls, and Hermes returns tool results in the next serialized transcript.

When bearer authentication is enabled, set the same token as model.api_key.

OpenClaw

Configure OpenClaw's openai-completions adapter in ~/.openclaw/openclaw.json, following the custom provider model. Set a bridge token in the environment used by both processes:

export FACTORY_DROID_OPENAI_API_KEY="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"

Add this JSON5 configuration:

{
  agents: {
    defaults: {
      model: { primary: "factory-droid/gpt-5.4" },
      models: {
        "factory-droid/gpt-5.4": {
          alias: "GPT-5.4 via Factory Droid",
        },
      },
    },
  },
  models: {
    mode: "merge",
    providers: {
      "factory-droid": {
        baseUrl: "http://127.0.0.1:8787/v1",
        apiKey: "${FACTORY_DROID_OPENAI_API_KEY}",
        api: "openai-completions",
        timeoutSeconds: 600,
        models: [
          {
            id: "gpt-5.4",
            name: "GPT-5.4 via Factory Droid",
            reasoning: true,
            input: ["text"],
            compat: {
              supportsDeveloperRole: true,
              supportsReasoningEffort: true,
              supportsUsageInStreaming: true,
              supportsTools: true,
              supportsStrictMode: false,
            },
          },
        ],
      },
    },
  },
}

Here, the first factory-droid segment names the OpenClaw provider. The gpt-5.4 segment and model id are the explicit Droid model ID.

Start the bridge from the working directory Droid should access, then restart the OpenClaw gateway. Verify the configuration with:

openclaw doctor
openclaw models list

OpenClaw retains ownership of tool execution. Tool calling, tool-result continuation, streaming usage, developer messages, and reasoning effort are enabled explicitly. Strict tool schemas, temperature control, multimodal SDK attachments, and parallel tool calls remain unsupported by this bridge.

VS Code

VS Code's Bring Your Own Key Custom Endpoint provider speaks Chat Completions, so it can drive the bridge without a Copilot plan or a GitHub sign-in.

Open the model picker, select Manage Language Models, then Add Models and Custom Endpoint. Pick Chat Completions as the API type. VS Code opens chatLanguageModels.json; add this provider entry:

[
  {
    "name": "Factory Droid",
    "vendor": "customendpoint",
    "apiKey": "none",
    "apiType": "chat-completions",
    "models": [
      {
        "id": "gpt-5.4",
        "name": "GPT-5.4 via Factory Droid",
        "url": "http://127.0.0.1:8787/v1/chat/completions",
        "toolCalling": true,
        "vision": true,
        "thinking": true,
        "streaming": true,
        "supportsReasoningEffort": ["low", "medium", "high", "xhigh"],
        "reasoningEffortFormat": "chat-completions",
        "maxInputTokens": 180000,
        "maxOutputTokens": 20000
      }
    ]
  }
]

A ready-to-use provider entry covering every model the reference host exposed is committed at examples/vscode/chatLanguageModels.json. Copy its contents into VS Code's chatLanguageModels.json, then adjust apiKey and the endpoint URL. Regenerate it against a running bridge, which reads image support and reasoning effort levels from GET /v1/models:

uv run python scripts/generate_vscode_models.py
uv run python scripts/generate_vscode_models.py --all-models
uv run python scripts/generate_vscode_models.py --model gpt-5.4 --model claude-opus-5
uv run python scripts/generate_vscode_models.py --all-models --verify

The committed example lists every model the reference host exposed. GET /v1/models returns the whole Droid catalog, including models an organization policy blocks for the signed-in account, so generate your own file with --verify: it starts a session for every selected model, drops the ones Droid refuses, and prints the refusals to stderr. No model turn runs, so verification costs no tokens; it spawns Droid directly and takes roughly a minute for a full catalog.

id is the explicit Droid model ID. Replace it with another ID from GET /v1/models or droid exec --help, such as gemini-3.1-pro-preview or the factory-droid alias. Each ID needs its own entry in models.

When bearer authentication is enabled, set the same token as apiKey. The placeholder none works only on a loopback bridge without a configured token. A mismatch makes VS Code report Invalid API key., which is the bridge answering 401; verify the token with curl -H "Authorization: Bearer $FACTORY_DROID_OPENAI_API_KEY" http://127.0.0.1:8787/v1/models.

id, name, url, toolCalling, vision, maxInputTokens, and maxOutputTokens are required by the model configuration reference. Omitting vision makes VS Code report a missing-property error for the entry.

toolCalling: true is required for the model to appear in agent sessions. vision: true enables image and PDF attachments, which the bridge forwards as base64 data: URIs; set it to false for text-only Droid models. thinking and supportsReasoningEffort surface the reasoning effort picker; the bridge maps those levels to Droid reasoning effort.

VS Code retains ownership of tool execution. The bridge converts Droid's strict tool protocol into OpenAI tool_calls, and VS Code returns tool results in the next serialized transcript.

BYOK covers chat and utility tasks only. Inline suggestions (code completions), semantic search, and embedding-dependent features still require a Copilot plan and a GitHub account. Without a GitHub sign-in, set chat.utilityModel and chat.utilitySmallModel to a configured model so title generation, commit messages, and intent detection keep working.

The bridge defaults to two concurrent Droid subprocesses (FACTORY_DROID_OPENAI_MAX_CONCURRENCY=2), so a chat turn and one utility task can run side by side. Raise that limit, or point utility tasks at a separate lightweight model, to avoid serial stalls. Utility tasks pay the same per-request cost as chat, so pointing chat.utilitySmallModel at a cheap model keeps title and commit-message generation out of the way.

Coverage with VS Code BYOK

What works through the Custom Endpoint provider:

• Chat and agent sessions, including tool calling
• Streaming responses
• Reasoning effort picker (thinking, supportsReasoningEffort)
• Structured output via response_format
• Inline image and PDF attachments sent as data: URIs

What VS Code BYOK does not cover, regardless of the bridge:

• Inline suggestions (code completions)
• Semantic search and embedding-dependent features
• Any feature that requires a Copilot plan or GitHub sign-in

What the bridge accepts but ignores when VS Code sends it: temperature, top_p, seed, max_tokens, max_completion_tokens, logprobs, and top_logprobs. These are documented in the API compatibility table. Remote http(s) image URLs are rejected with 400; send vision inputs as base64 data: URIs instead. VSCode does not auto-discover models from a Custom Endpoint, so every Droid model ID you want in the picker needs its own entry in chatLanguageModels.json.

API reference

Endpoints

Method	Endpoint	Description
GET	/health	Process health check
GET	/v1/models	Bridge alias and dynamically discovered Droid models
POST	/v1/chat/completions	Chat completions and streaming
GET	/v1/factory/sessions/{id}/context	Context statistics and breakdown
POST	/v1/factory/sessions/{id}/compact	Compact history into a new session
POST	/v1/factory/sessions/{id}/fork	Fork a session with context preserved
PATCH	/v1/factory/sessions/{id}	Rename a session
DELETE	/v1/factory/sessions/{id}	Close a session
GET	/metrics	Prometheus-style bridge metrics
GET	/openapi.json	OpenAPI 3.1 contract
GET	/docs	Interactive Swagger UI
GET	/redoc	Interactive ReDoc reference

OpenAPI contract

The versioned OpenAPI 3.1 document describes the supported HTTP contract. Regenerate and validate it after changing routes or models:

uv run python scripts/generate_openapi.py
uv run openapi-spec-validator openapi.json

Tests compare the committed document with FastAPI's generated schema. CI also validates it independently with openapi-spec-validator, so malformed or stale API contracts fail automatically. The contract covers the implemented OpenAI-compatible subset, not the complete OpenAI platform. Separate contract tests exercise models, non-streaming chat, streaming chat, usage, and function tool calls through the official openai Python client.

Request fields

Field	Support	Behavior
model	Yes	Alias uses Droid default; other values become Droid model IDs
messages	Yes	Complete transcript serialized in original order
tools	Yes	OpenAI function tools
tool_choice	Yes	auto, none, required, or one named function
stream	Yes	OpenAI-compatible SSE chunks
stream_options.include_usage	Yes	Emits null usage on normal chunks and one final usage-only chunk
stream_options.include_obfuscation	No	Accepted but ignored
n	Yes	Runs n Droid turns sequentially, capped by the server
stop	Yes	String or list; truncates output and interrupts the turn
parallel_tool_calls	Yes	true allows several tool calls per turn, false allows one
factory_droid_session_id	Yes	Bridge-specific session continuation, disabled by default
factory_droid_status	Yes	Bridge-specific Droid working-state SSE events
reasoning_effort	Yes	Mapped to Droid reasoning effort
factory_droid_reasoning_effort	Yes	Bridge-specific override
timeout	Yes	Per-request value capped by server timeout
temperature, top_p, penalties, seed	No	Accepted but ignored
max_tokens, max_completion_tokens	No	Accepted but ignored
response_format	Yes	json_schema and json_object; validated again before completion
functions, function_call	No	Legacy function-calling fields are ignored
modalities, audio	No	Accepted but ignored
logprobs, top_logprobs, logit_bias	No	Accepted but ignored
store, metadata, user, safety_identifier	No	Accepted but ignored
prompt_cache_key, cache options and retention	No	Accepted but ignored
prediction, service_tier, verbosity	No	Accepted but ignored
web_search_options	No	Accepted but ignored
Unknown fields	Accepted	Ignored by the bridge

The bridge currently returns one choice with index 0.

Model selection

factory-droid is a server-side alias. It passes no explicit model ID and lets Droid use its configured default model.

Any other model value is forwarded as model_id during Droid session initialization:

{
  "model": "gpt-5.4",
  "messages": [{"role": "user", "content": "Hello"}]
}

For the official OpenAI client, the same explicit selection is:

client.chat.completions.create(
    model="gemini-3.1-pro-preview",
    messages=[{"role": "user", "content": "Hello"}],
)

Call GET /v1/models to list the bridge alias and the models currently available to the authenticated Droid CLI. Factory-specific metadata includes the display name, reasoning efforts, and image and PDF support.

Discovery starts a short-lived Droid session, so the catalog is cached for FACTORY_DROID_OPENAI_MODEL_CACHE_SECONDS and concurrent callers share a single discovery attempt. If discovery cannot start or authenticate Droid, the endpoint preserves compatibility by serving the last known catalog, or the bridge alias alone, and marks the response with the x-factory-droid-model-discovery: degraded header while incrementing factory_droid_openai_model_discovery_failures_total.

The Droid catalog also lists models an organization policy blocks for the signed-in account. Droid rejects those when a session starts, which the bridge maps to 404 model_not_found and remembers for FACTORY_DROID_OPENAI_MODEL_QUARANTINE_SECONDS. While a model is quarantined, GET /v1/models withholds it and reports how many entries were withheld in the x-factory-droid-models-quarantined header, chat requests for it are refused without starting Droid, and the warm-session pool stops warming it. Set the value to 0 to disable quarantining.

Reasoning

Supported values follow droid-sdk-python:

• none
• dynamic
• off
• minimal
• low
• medium
• high
• xhigh
• max

Reasoning text appears as both reasoning and reasoning_content for broad client compatibility.

Multimodal attachments

Image and file content parts are delivered over the Droid SDK's native attachment channel instead of being inlined into the prompt text. The transcript keeps a short placeholder in their place, so positional context survives without sending the payload twice.

client.chat.completions.create(
    model="gpt-5.4",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What is in this screenshot?"},
                {
                    "type": "image_url",
                    "image_url": {"url": "data:image/png;base64,<base64>"},
                },
            ],
        }
    ],
)

Supported inline media types are image/jpeg, image/png, image/gif, image/webp, application/pdf, and text/plain. Remote http(s) URLs and file_id references are rejected with HTTP 400: the bridge never fetches remote content on a client's behalf. Model support for images and PDFs still depends on the selected Droid model.

Session continuity

By default every request creates a new Droid session and serializes the whole transcript. Setting FACTORY_DROID_OPENAI_SESSION_CONTINUITY=true lets a client continue an existing session instead, so only the new turn is sent:

first = client.chat.completions.create(model="gpt-5.4", messages=messages)
session_id = first.model_extra["factory_droid_session_id"]

second = client.chat.completions.create(
    model="gpt-5.4",
    messages=[*messages, assistant_reply, next_question],
    extra_body={"factory_droid_session_id": session_id},
)

The session ID is also returned in the x-factory-droid-session-id response header, and announced in the first streaming chunk.

Only sessions this bridge process created can be continued. Session IDs stored by the local Droid CLI - including your own interactive sessions - are rejected with HTTP 404, so a client cannot read back conversations it does not own. Restarting the bridge clears the set of continuable sessions.

The same guard applies to the Factory session extension endpoints. They are available only when continuity is enabled and only for IDs created by the current bridge process:

curl http://127.0.0.1:8787/v1/factory/sessions/$SESSION_ID/context

curl -X POST \
  http://127.0.0.1:8787/v1/factory/sessions/$SESSION_ID/compact \
  -H 'Content-Type: application/json' \
  -d '{"custom_instructions":"Keep API decisions."}'

curl -X POST \
  http://127.0.0.1:8787/v1/factory/sessions/$SESSION_ID/fork

curl -X PATCH \
  http://127.0.0.1:8787/v1/factory/sessions/$SESSION_ID \
  -H 'Content-Type: application/json' \
  -d '{"title":"API design"}'

curl -X DELETE \
  http://127.0.0.1:8787/v1/factory/sessions/$SESSION_ID

Compaction and fork return a new session_id, which the bridge registers for later continuation. Context responses contain aggregate token categories, not message or prompt content. Context, fork, rename, and close run no model turn, so they leave the session's tool settings untouched; compaction runs a turn and therefore disables the tool catalog first.

Droid status events

factory_droid_status: true adds Droid working-state transitions to a streaming response as an extra factory_droid_status key inside the chunk delta:

{"choices":[{"index":0,"delta":{"factory_droid_status":"executing_tool"}}]}

States follow the SDK: idle, streaming_assistant_message, waiting_for_tool_confirmation, executing_tool, and compacting_conversation. The field is absent unless explicitly requested, so strict OpenAI clients are unaffected.

Errors

Non-streaming failures use the OpenAI error object:

{
  "error": {
    "message": "Factory Droid executable was not found: droid",
    "type": "factory_droid_unavailable",
    "param": null,
    "code": null
  }
}

HTTP status	Meaning
400	Invalid request or unsupported option
401	Missing or invalid bearer token
404	Unknown Factory Droid session
413	Request body or serialized transcript over the configured limit
429	Droid request queue full, with a Retry-After header
502	Droid SDK, process, or bridge protocol failure
503	Droid executable unavailable
504	Request timeout

After streaming headers are sent, errors arrive as an SSE error object followed by [DONE].

Configuration

Environment variable	Default	Description
FACTORY_DROID_OPENAI_HOST	127.0.0.1	Listen address
FACTORY_DROID_OPENAI_PORT	8787	Listen port
FACTORY_DROID_OPENAI_API_KEY	unset	Optional bearer token
FACTORY_DROID_PATH	droid	Droid executable path
FACTORY_DROID_OPENAI_WORKDIR	current directory	Droid working directory
FACTORY_DROID_OPENAI_TIMEOUT_SECONDS	600	Maximum request duration
FACTORY_DROID_OPENAI_BODY_TIMEOUT_SECONDS	30	Maximum time to receive a request body
FACTORY_DROID_OPENAI_MAX_CONCURRENCY	2	Concurrent Droid subprocesses
FACTORY_DROID_OPENAI_WARM_SESSIONS	MAX_CONCURRENCY + 1	Pre-started Droid sessions kept ready (0 disables)
FACTORY_DROID_OPENAI_WARM_SESSION_TTL_SECONDS	600	Age at which an unused warm session is replaced
FACTORY_DROID_OPENAI_DETACHED_CLEANUP	true	Tear down Droid processes after the response is sent
FACTORY_DROID_OPENAI_MAX_QUEUE_SIZE	8	Requests waiting for a Droid slot
FACTORY_DROID_OPENAI_RETRY_AFTER_SECONDS	1	Retry-After value sent with 429
FACTORY_DROID_OPENAI_MAX_REQUEST_BYTES	4194304	Request body size limit
FACTORY_DROID_OPENAI_MAX_MESSAGES	512	Messages accepted per request
FACTORY_DROID_OPENAI_MAX_TOOLS	128	Tool schemas accepted per request
FACTORY_DROID_OPENAI_MAX_TRANSCRIPT_BYTES	4194304	Serialized transcript size limit
FACTORY_DROID_OPENAI_MAX_TOOL_SCHEMA_BYTES	1048576	Serialized tool schema size limit
FACTORY_DROID_OPENAI_MAX_STRUCTURED_OUTPUT_BYTES	1048576	Buffered structured output size limit
FACTORY_DROID_OPENAI_MAX_JSON_DEPTH	32	Request JSON nesting limit
FACTORY_DROID_OPENAI_PROCESS_GRACE_SECONDS	2	Wait before killing a Droid process
FACTORY_DROID_OPENAI_CLEANUP_TIMEOUT_SECONDS	4	Total budget for session cleanup
FACTORY_DROID_OPENAI_UVICORN_LIMIT_CONCURRENCY	64	Uvicorn connection limit
FACTORY_DROID_OPENAI_UVICORN_BACKLOG	128	Uvicorn listen backlog
FACTORY_DROID_OPENAI_MAX_TOOL_CALLS	8	Tool calls accepted per Droid turn
FACTORY_DROID_OPENAI_MAX_ATTACHMENTS	16	Inline attachments per request
FACTORY_DROID_OPENAI_MAX_ATTACHMENT_BYTES	8388608	Total encoded attachment bytes
FACTORY_DROID_OPENAI_MAX_CHOICES	4	Upper bound for n
FACTORY_DROID_OPENAI_MAX_STOP_SEQUENCES	4	Upper bound for stop entries
FACTORY_DROID_OPENAI_SESSION_CONTINUITY	false	Allow continuing bridge-created sessions
FACTORY_DROID_OPENAI_MAX_TRACKED_SESSIONS	256	Continuable sessions kept in memory
FACTORY_DROID_OPENAI_WORKTREE	unset	Run Droid in a git worktree
FACTORY_DROID_OPENAI_APPEND_SYSTEM_PROMPT_FILE	unset	File appended to the Droid system prompt
FACTORY_DROID_OPENAI_MODEL_ALIAS	factory-droid	Alias using Droid default model
FACTORY_DROID_OPENAI_MODEL_CACHE_SECONDS	300	GET /v1/models catalog cache lifetime
FACTORY_DROID_OPENAI_MODEL_QUARANTINE_SECONDS	900	How long a refused model stays withheld
FACTORY_DROID_OPENAI_MCP_SETTLE_SECONDS	0	MCP initialization window before tool discovery
FACTORY_DROID_OPENAI_LOG_LEVEL	info	Bridge and Uvicorn log level
FACTORY_DROID_OPENAI_LOG_FORMAT	text	Bridge log rendering: text or json

Command-line options:

usage: factory-droid-openai [-h] [--host HOST] [--port PORT]
                            [--limit-concurrency LIMIT_CONCURRENCY]
                            [--backlog BACKLOG]
                            [--log-level {critical,error,warning,info,debug,trace}]
                            [--log-format {text,json}]
                            [--access-log | --no-access-log]

Environment variables define defaults. The command-line options override the matching server settings.

Limits and metrics

Every request passes two gates before a Droid process is involved.

Size limits are enforced while the body is still arriving, so an oversized payload is rejected with 413 instead of being buffered in full. Body bytes, message count, tool count, serialized transcript size, tool schema size, and JSON nesting depth each have their own bound. Body size, nesting depth, the body timeout, and request metrics cover every HTTP route, including GET /v1/models and the /v1/factory/sessions endpoints.

Admission control then bounds how many requests reach Droid. Requests beyond FACTORY_DROID_OPENAI_MAX_CONCURRENCY wait in a queue of at most FACTORY_DROID_OPENAI_MAX_QUEUE_SIZE; once that is full the bridge answers 429 with a Retry-After header rather than queueing without limit.

GET /metrics renders Prometheus-style text. It is excluded from the OpenAPI contract because it is an operational endpoint, not part of the OpenAI API surface.

Metric	Meaning
factory_droid_openai_requests_total	Requests by outcome
factory_droid_openai_request_duration_seconds	Request duration sum and count
factory_droid_openai_queue_wait_seconds	Time spent waiting for admission
factory_droid_openai_droid_startup_seconds	Droid session startup time
factory_droid_openai_ttft_seconds	Time to first token
factory_droid_openai_active_sessions	Droid sessions running now
factory_droid_openai_queued_requests	Requests waiting for admission
factory_droid_openai_overload_rejections_total	Requests rejected with 429
factory_droid_openai_payload_rejections_total	Requests rejected with 413
factory_droid_openai_forced_kills_total	Droid processes that needed a kill
factory_droid_openai_model_discovery_failures_total	Failed Droid model discovery attempts
factory_droid_openai_model_quarantines_total	Models withheld after Droid refused them
factory_droid_openai_warm_sessions	Warm Droid sessions ready now
factory_droid_openai_warm_session_hits_total	Requests served from a warm session
factory_droid_openai_warm_session_retunes_total	Warm sessions repointed at another model
factory_droid_openai_warm_session_misses_total	Requests that had to start their own session
factory_droid_openai_warm_session_failures_total	Failed warm-up attempts
factory_droid_openai_pending_reaps	Droid teardowns still running in the background

forced_kills_total counts processes that had to be killed with a signal because they did not exit within FACTORY_DROID_OPENAI_PROCESS_GRACE_SECONDS. droid exec shuts down cleanly in about a second, so a rising counter means stuck processes, not ordinary traffic.

Serve /metrics on a loopback interface or behind your own access control; it carries no prompt content, but it does expose traffic shape.

Warm sessions

Starting a Droid session and tearing it down again dominates the cost of a short request. Measured on an Apple M-series laptop with a trivial prompt:

Phase	Cost	On the request path?
droid exec spawn and JSON-RPC connect	3 ms	yes
droid.initialize_session	2.4-3.2 s	no, pre-warmed
Disabling Factory-native tools	6 ms	no, pre-warmed
Prompt serialization and send	4 ms	yes
Model time to first token	2.8-3.8 s	yes
Session close and process exit	1.0 s	no, detached

So the bridge keeps FACTORY_DROID_OPENAI_WARM_SESSIONS sessions initialized and idle, and hands one to each incoming request. A warm session serves a single turn and is then discarded, so requests stay isolated exactly as before. End to end, the same prompt drops from about 7.2 s to about 2.8 s, leaving little more than model time. The default keeps one spare session per concurrency slot, so back-to-back requests still find one warm while the pool refills.

Warm sessions are keyed by the model and reasoning effort they were initialized with. An exact match serves a turn with no extra round trip. Otherwise the bridge repoints a session warmed for another model at the requested one, which takes 4-18 ms instead of a 2.4-3.2 s startup, and logs pool.retune plus droid.session_retuned. Requests that ask for the model alias, or for the model's default reasoning effort on a session that carries an explicit one, cannot be expressed as an update and still log pool.miss. The pool tracks the settings recent traffic used, so repeat traffic converges on exact matches.

Teardown runs after the response is finished, so session close and the second droid exec needs to exit no longer delay the last token. Set FACTORY_DROID_OPENAI_DETACHED_CLEANUP=false to wait for teardown on the request path instead, and FACTORY_DROID_OPENAI_WARM_SESSIONS=0 to disable pre-warming. Either way one Droid process serves one request; warm sessions only move the startup cost off the request path, and each idle session holds a Droid process while it waits.

Logging

The bridge logs structured events on top of Uvicorn's access log. Every line carries an event name, the OpenAI request id, and phase timings in milliseconds, so a slow turn can be attributed to a phase instead of guessed at.

factory-droid-openai --log-level debug
factory-droid-openai --log-level info --log-format json
factory-droid-openai --log-level debug --no-access-log

Level	Content
warning	Rejections, failures, degraded model discovery, quarantined models, forced process kills
info	One chat.completed summary per request with phase timings and token usage
debug	Per-phase events: prompt built, admission, warm-session hit, retune or miss, Droid startup, session ready, first token, turn complete, cleanup
trace	One line per Droid SDK event kind

DEBUG    2026-07-27T14:10:02+0200 event=chat.received request_id=chatcmpl-8f2a model=gpt-5.4 stream=true messages=6 tools=14
DEBUG    2026-07-27T14:10:02+0200 event=chat.prompt_built request_id=chatcmpl-8f2a prompt_bytes=48213 allowed_tools=14 prompt_ms=7.9
DEBUG    2026-07-27T14:10:02+0200 event=chat.admitted request_id=chatcmpl-8f2a queue_ms=0.2
DEBUG    2026-07-27T14:10:02+0200 event=pool.hit request_id=chatcmpl-8f2a model=gpt-5.4 warm_sessions=1
DEBUG    2026-07-27T14:10:02+0200 event=droid.session_ready request_id=chatcmpl-8f2a warm=true session_ready_ms=3.1
DEBUG    2026-07-27T14:10:06+0200 event=chat.first_token request_id=chatcmpl-8f2a ttft_ms=3908.4
INFO     2026-07-27T14:10:11+0200 event=chat.completed request_id=chatcmpl-8f2a outcome=success stream=true input_tokens=9123 output_tokens=412 total_ms=8712.6
DEBUG    2026-07-27T14:10:12+0200 event=droid.cleanup request_id=chatcmpl-8f2a cleanup_ms=1021.4

droid.connected and droid_startup_ms only appear on a cold session, pool.retune and droid.session_retuned only when a session warmed for another model is reused, and cleanup is logged after the response because it is detached.

Phase fields, in the order they occur:

Field	Covers
prompt_ms	Transcript and tool-schema serialization in the bridge
queue_ms	Waiting for a Droid concurrency slot
droid_startup_ms	droid exec process spawn and SDK handshake
retune_ms	Repointing a warm session at the requested model
session_ready_ms	Elapsed until the Droid session accepts a prompt
prompt_sent_ms	Elapsed until the prompt was handed to Droid
ttft_ms	Elapsed until the first text or reasoning delta
turn_ms	Elapsed until Droid reported turn completion
total_ms	Whole request, including response encoding

One Droid process serves each request. A warm session brings session_ready_ms down to a few milliseconds, so a request that logs seconds there ran on a cold session; check for a preceding pool.miss. A large queue_ms means requests are serialized behind FACTORY_DROID_OPENAI_MAX_CONCURRENCY, which also covers VS Code utility tasks such as title generation.

Log lines carry no prompt or completion text: sizes, counts, timings, model IDs, and session IDs only.

Authentication

Local loopback use requires no bridge token by default. OpenAI clients still expect an API key value, so use a placeholder such as none.

Enable bearer authentication:

export FACTORY_DROID_OPENAI_API_KEY="$(python -c 'import secrets; print(secrets.token_urlsafe(32))')"
factory-droid-openai

Send the token:

curl --fail http://127.0.0.1:8787/v1/models \
  -H "Authorization: Bearer $FACTORY_DROID_OPENAI_API_KEY"

Never expose an unauthenticated bridge on a non-loopback interface. See SECURITY.md before remote deployment.

Tool execution safety

The OpenAI client, not Factory Droid, owns tool execution.

For every Droid session, the bridge:

1. Selects Droid's auto interaction mode with autonomy off.
2. Requests no additional MCP servers from the SDK.
3. Cancels permission requests and interactive questions.
4. Optionally waits FACTORY_DROID_OPENAI_MCP_SETTLE_SECONDS for configured MCP servers, which only widens the verification snapshot and is off by default.
5. Discovers the native and MCP tool catalog known at that moment through droid.list_tools.
6. Disables every discovered tool ID and verifies no unexpected tool remains. Tools that appear after this point stay denied, because Droid defaults to deny once a disabled set has been sent.
7. Rejects any Factory-native tool event as a second line of defense.
8. Validates generated tool names against the request schema.
9. Requires tool arguments to be a JSON object with unique keys.
10. Caps the number of tool calls accepted in one turn.
11. Rejects any non-whitespace output between or after tool calls.
12. Refuses to fetch remote attachment URLs.
13. Only continues Droid sessions this bridge process created.

Set FACTORY_DROID_OPENAI_WORKDIR to the smallest directory required by your workflow.

Bridge architecture limits

This is a compatibility bridge, not a native OpenAI inference implementation.

Limitation	Effect
No native SDK chat-history input	History is serialized into one prompt unless session continuity is enabled
No native SDK external-tool schemas	Tool definitions are serialized into the prompt
No native SDK tool-result continuation	Each request runs a fresh Droid session; warm sessions remove its startup cost, not its empty state
Python SDK protocol drift	A small isolated compatibility shim supplies structured output, tool controls, and session RPCs
Session-per-request execution	Prompt caching differs from a native inference endpoint
Strict tool marker protocol	Invalid generated tool payloads fail closed
Sequential tool calls only	Calls arrive one after another, never concurrently

Troubleshooting

Droid executable not found

factory_droid_unavailable

Set the executable path:

export FACTORY_DROID_PATH="$HOME/.local/bin/droid"

Authentication failure

An HTTP 401 means FACTORY_DROID_OPENAI_API_KEY is configured and the request token is missing or incorrect. This is separate from Factory account authentication used by the Droid CLI.

Factory-native tool blocked

factory_native_tool_blocked

Droid attempted to use one of its own tools instead of returning bridge protocol output. Retry with a clearer request or inspect the supplied system and user messages for conflicting instructions.

Incomplete response

factory_incomplete_response

The Droid stream ended without a turn-complete event. Check Droid CLI health, request timeout, and bridge logs.

One slow request after switching models

The warm pool holds sessions for the model and reasoning effort recent traffic used. A session warmed for another model is repointed in milliseconds and logs pool.retune, so only requests that cannot be expressed as a settings update - the model alias, or a model default effort on a session with an explicit one - pay full startup and log pool.miss with a multi-second session_ready_ms. Raise FACTORY_DROID_OPENAI_WARM_SESSIONS when clients alternate between models faster than the pool refills.

Every request logs a forced kill

Not expected any more. droid exec exits cleanly in about a second, so raise FACTORY_DROID_OPENAI_PROCESS_GRACE_SECONDS if teardown is being cut short on a slower host; FACTORY_DROID_OPENAI_CLEANUP_TIMEOUT_SECONDS must stay above it. With detached cleanup the kill happens after the response, so it costs the client nothing either way.

Model not allowed by organization policy

Droid lists the full catalog but refuses to start a session for models the account's organization policy blocks. The bridge answers 404 model_not_found, logs models.quarantined, and withholds the model from GET /v1/models for FACTORY_DROID_OPENAI_MODEL_QUARANTINE_SECONDS, so a client that rebuilds its model list recovers on its own. Regenerate client configuration with uv run python scripts/generate_vscode_models.py --all-models --verify to drop blocked models up front.

Development

Install locked development dependencies:

uv sync --all-extras

Run required validation:

uv run ruff check .
uv run ruff format --check .
uv run mypy
uv run python scripts/generate_openapi.py
uv run openapi-spec-validator openapi.json
uv run pytest --cov=factory_droid_openai --cov-report=term-missing
uv build

Automated test layers

Layer	CI status	Coverage
Unit tests	Automated	Configuration, protocol parsing, runner lifecycle, CLI, and response mapping
HTTP integration	Automated	Authentication, JSON, SSE, errors, timeouts, and disconnect cleanup through ASGI
OpenAI client contract	Automated	Official openai client models, chat, streaming, usage, and function tools
OpenAPI contract	Automated	Generated-schema drift plus independent OpenAPI 3.1 validation
Package smoke test	Automated	Build, isolated wheel install, CLI, live process health, models, schema, and validation error
Live Factory inference	Manual	Requires an authenticated local Droid CLI and can consume account quota

The automated suite does not access Factory services or require credentials. Public CI intentionally excludes live inference because it requires a Factory account. The local curl and client examples above provide the credentialed end-to-end smoke path before deployment.

See CONTRIBUTING.md for contribution requirements and SECURITY.md for private vulnerability reporting. Release history is published on the releases page and generated from the commit history by Release Please.

License

Licensed under the Apache License 2.0.