AI harness software is what turns a language model into a working agent. Droid is the coding agent harness from Factory (legally: The San Francisco AI Factory Inc.), connecting language models with tools, MCP servers, and a risk-based approval system, and running in the terminal, as a desktop app, and in the cloud. As of 20 September 2026.
What is Droid?
Droid is Factory’s coding agent and, according to the documentation, “runs the same everywhere”: as a terminal CLI, as a desktop app, or in the cloud via the web platform. Depending on the platform, the CLI is installed via install script (curl -fsSL https://app.factory.ai/cli | sh), Homebrew (brew install --cask droid), or npm (npm install -g droid); on Linux, the xdg-utils package is additionally required. After installation, you navigate to a project directory and start an interactive session with the droid command, authenticating via browser login.
Droid as an AI harness
Droid’s agent loop runs either interactively in a terminal interface with visible approvals and diffs, or non-interactively via the headless command droid exec, which carries a task through to completion in a single pass and returns text, JSON, or stream-jsonrpc output. Project context comes from an AGENTS.md file covering architecture, build/test commands, and conventions. Permissions are governed by the so-called Autonomy Level across four tiers – Off, Low, Medium, High – which determines, based on risk, which commands and MCP tools run without asking; a built-in blocklist prevents certain destructive commands regardless of the chosen level, and separate sandbox checks for file and network access can force additional approvals. A /rewind menu, or the /rewind-conversation command, lets you undo changes made during a session. hooks.json, SKILL.md files under .factory/skills/, and a plugin system (.factory-plugin/) let you extend Droid with deterministic automation, reusable workflows, and extensions distributed across a team. For scoped subtasks, custom subagents (“Custom Droids”) can be defined as Markdown files with their own system prompt, model, and tool set, and invoked via the Task tool in the foreground or background.
Models and model access
Droid does not name a single fixed default model; instead, it connects via built-in connections as well as via Custom Models (BYOK) to your own Claude- or GPT-compatible API keys, or to routers such as OpenRouter or local servers such as Ollama and vLLM. This is configured in the settings.json file under the customModels key, with the fields model, baseUrl, provider, and optionally apiKeyHelper. According to the documentation, this custom-endpoint connection applies to the Droid CLI and desktop app, not to the hosted web/mobile platform. Whether this configurable endpoint can specifically be routed through the AI Gateway should be checked case by case against the gateway’s interface.
Enterprise use: data protection and governance
For Enterprise customers with European data-residency requirements, Factory offers an EU organization with its own API backend and web app hosted within the European Economic Area; organization and billing data remain in the United States according to the documentation, with standard contractual clauses intended to cover transfers there. Droid can also run fully air-gapped, with custom certificate authorities, mutual TLS, and centrally distributed admin policies for model access, network rules, and IP restrictions. Important for enterprises: source code and prompts go to whichever model provider is configured and should be reflected in the records of processing activities beforehand.
Orchestrating Droid with herdr
herdr lists Droid in its agent table with the integration level “screen manifest” and “session”. This means herdr detects Droid sessions via the terminal screen manifest and additionally folds them into its shared session management, for example to start and observe several agents in a coordinated way.
Where it fits: developer harness and enterprise harness
Droid is a tool for developers working in the terminal, on the desktop, or in the cloud. For non-developer staff, CompanyGPT is the matching harness: the same underlying idea of tools, context, and permissions, but built for enterprise roles, integrations, and audit requirements. The AI Gateway can serve as a shared cost and governance layer across different harnesses. For an overview of other coding agents and orchestrators, see our overview of all AI harnesses. We help companies introduce coding harnesses like Droid safely and connect them to their existing AI governance.