The shape
Bottom line
Use Droid to delegate work, not to sit next to it.
Factory Droid is the strongest fit for teams that already live in GitHub, Linear, and Jira and want to fan out queued tickets across many repos with model-agnostic agents that can also run inside customer infrastructure. It is a weaker fit for solo developers who want the most polished interactive pair-programmer, where Claude Code or Codex CLI still feel tighter. The most common 2026 pattern is to keep Claude Code or Codex local and use Droid for asynchronous, ticket-style delegation.
Snapshot
What Droid is
| Axis | Factory Droid |
|---|---|
| Product shape | Multi-surface software-development agent: terminal CLI, headless `droid exec`, IDE plugins, desktop app, web/mobile, Slack/Teams, and integrations with Linear, Jira, and GitHub. |
| Install | `curl -fsSL https://app.factory.ai/cli | sh`, `irm https://app.factory.ai/cli/windows | iex`, or `npm install -g droid`. Run `droid` in any project. |
| Default posture | Read-only by default in headless mode; mutating work requires `--auto low|medium|high` or the explicit `--skip-permissions-unsafe` bypass. |
| Models | Anthropic and OpenAI as fully tested first-party providers; BYOK across OpenAI-compatible gateways such as OpenRouter, Fireworks, Together AI, Ollama, vLLM, Baseten, DeepInfra, Google Gemini, Groq, and Hugging Face. |
| Extensibility | Custom Droids (subagents in `.factory/droids/`), Skills (`.factory/skills/`), Hooks, MCP servers, AGENTS.md context, and plugins. |
| Enterprise | Three documented modes: cloud-managed, hybrid in-customer-infra, and fully airgapped. Data plane stays with the customer; LLM traffic routes through customer gateways. |
| Pricing | Pro $20/mo, Plus $100/mo, Max $200/mo, Teams custom up to 150 seats, Enterprise custom with on-prem and dedicated compute. |
Where Droid runs
Five surfaces, one agent
| Surface | What it does | Best when |
|---|---|---|
| Droid CLI (`droid`) | Interactive terminal session with diff viewer and approvals; bash mode toggle with `!`. | Local pair-programming and exploratory work on your own machine. |
| Droid Exec (`droid exec`) | One-shot non-interactive run that writes to stdout/stderr with text, JSON, or stream-json output. | CI/CD pipelines, scheduled sweeps, and shell-parallel batch work. |
| Desktop App | Local desktop client that connects Droid to your machine. | You want a GUI surface without leaving the local filesystem. |
| Web and mobile | Hosted Droid runs in a browser with no local setup, plus mobile views. | Kicking off or monitoring a run away from your dev machine — though continuity between local and mobile is reportedly incomplete. |
| IDE, Slack, and PM | VS Code, JetBrains, and Zed plugins; Slack/Teams chat; Linear/Jira issue assignment and mentions. | Team workflows where tickets, chat, and PRs already drive the work. |
Autonomy tiers
Permission gates that scale with risk
| Mode | Allows | Blocks | Use it for |
|---|---|---|---|
| Default (read-only) | File reads, `git status`, directory listing, planning. | All file edits, package installs, network mutations, git pushes. | Reviewing planned changes safely or generating reports. |
| `--auto low` | Basic file create/edit inside the project directory, doc updates, formatters. | System modifications and package installations. | Documentation, comments, and small in-tree edits. |
| `--auto medium` | Local package installs, network requests to trusted endpoints, `git commit`/`git pull`. | Git pushes, production changes. | Dev workflows where you want to install deps and run tests but stop short of pushing. |
| `--auto high` | Arbitrary code execution, git push, production deploys. | `sudo rm -rf /` and system-wide changes. | Trusted, well-scoped pipelines with strong observability. |
| `--skip-permissions-unsafe` | Everything, no confirmations. | Nothing. | Throwaway VMs or Docker sandboxes only — explicitly flagged as unsafe in the docs. |
Building blocks
Missions, Custom Droids, Skills, Hooks, MCP
| Primitive | What it is | Why it matters |
|---|---|---|
| Missions | Plan-and-execute multi-step workflows for larger projects. | Lets a single delegated ticket plan, branch, and execute several phases without re-prompting. |
| Custom Droids | Reusable subagent definitions in Markdown with YAML frontmatter at `.factory/droids/` (project) or `~/.factory/droids/` (personal). | Encodes role-specialized checklists — code, review, docs, test — with their own tools and model preferences. |
| Skills | Reusable capabilities at `.factory/skills/<name>/SKILL.md` or `skill.mdx`, optionally invocable via `/skill-name`. | Packages how work should be done independent of which Droid is running it; legacy `.agent/skills/` is also supported. |
| Hooks | Pre/post tool, session, and prompt event hooks. | Block dangerous commands, force validation, or wire in deterministic guardrails before model decisions. |
| MCP | Org allow/deny lists with project- and user-scoped MCP server configuration. | Lets enterprise teams curate which external context and tools an agent can reach. |
| Droid Computers / BYOM | Persistent local or cloud machines registered as Droid Computers; BYOM lets you register Linux, macOS, or Windows machines that connect outbound to `relay.factory.ai`. | No inbound ports needed, but the relay still depends on Factory infrastructure unless you are on the airgapped enterprise path. |
Enterprise deployment
Three modes; the data plane stays with you
| Mode | What runs where | When to choose it |
|---|---|---|
| Cloud-managed | Droid runs on developer machines and CI; Factory cloud handles orchestration. LLM traffic can still go through customer gateways. | You want the fastest path and accept Factory cloud as the control plane. |
| Hybrid enterprise | Droid processes run on customer VMs, containers, CI runners, and remote dev environments. LLM traffic routes through customer gateways; OTEL telemetry to customer collectors. Factory cloud may see only limited metadata. | Network segmentation, central governance, and central LLM gateways are mandatory. |
| Fully airgapped | No outbound connectivity to Factory cloud. Models served from on-prem or in-network endpoints; OTEL stays inside. | Classified, regulated, or fully isolated environments. |
Models and BYOK
Model-agnostic by design
Factory positions Droid as not locked into a single AI provider. Anthropic and OpenAI models are the fully tested first-party path, including current Claude (Opus and Sonnet) and OpenAI (GPT-5 and GPT-5-Codex) families. BYOK adds OpenAI-compatible gateways such as OpenRouter, Fireworks, Together AI, Ollama, vLLM, Baseten, DeepInfra, Google Gemini, Groq, and Hugging Face. API keys remain local and are not uploaded to Factory servers, but custom models work only in the CLI — not on the web or mobile clients.
Upgrade / downgrade
Choose by where you want the agent to live
| Pick | What you gain | What you lose | Choose it when |
|---|---|---|---|
| Simpler interactive Claude Code | Polished local pair-programmer with strong first-party Anthropic integration and tight diff loops. | Single-vendor model lock-in and weaker queued ticket delegation. | You're a solo developer who wants the cleanest interactive coding loop. |
| Lateral interactive Codex CLI | OpenAI-side polish with app, CLI, IDE, and automation surfaces. | Anthropic-only vs OpenAI-only choice; not model-agnostic. | You live inside the OpenAI ecosystem and want first-party tooling. |
| Target Factory Droid | Multi-surface delegation, BYOK across providers, Custom Droids/Skills/Hooks, hybrid and airgapped enterprise deployment, and the option to run on customer infrastructure. | Less polished interactive feel; background work surfaces problems at PR review; closed runtime. | Your team needs queued ticket execution across many repos and surfaces. |
| Lateral team-native GitHub Copilot coding agent | Issue-to-PR automation living entirely inside GitHub with ephemeral Actions-backed environments. | GitHub-only; less local interactive use; fewer model choices. | Your team already routes work through GitHub issues and PRs. |
| Lateral background Cursor background agents | IDE-native background runs with the editor as the primary surface. | Cursor lock-in; less of a multi-surface platform. | You prefer agents attached to your editor and don't need org-scale governance. |
| Upgrade for ownership OpenHands | Open-source agent runtime, self-hostable Docker/Kubernetes sandboxes, SDK, and remote agent server. | More platform engineering; less polished out of the box. | Long-term runtime ownership and inspectable code matter more than vendor polish. |
| Upgrade for managed work Devin | Most managed autonomous-engineering experience for backlog delegation. | Higher per-task cost; vendor-controlled environment. | You want a Devin-like experience and Devin-like budget. |
Reviews and user signal
What independent reviewers and benchmarks say
| Source | Signal type | Headline finding |
|---|---|---|
| Factory Terminal-Bench post (Sep 25, 2025) | Vendor benchmark | Droid set a 58.75% state-of-the-art on Terminal-Bench, with Opus 4.1 at 58.8%, GPT-5 at 52.5%, and Sonnet 4 at 50.5%; vendor-run, replicate before quoting. |
| Every — Vibe Check (Oct 30, 2025; updated Jan 17, 2026) | Pro review | Author canceled both Claude and ChatGPT Max plans for Droid, citing model switching, more deliberate behavior, and stronger plan mode; flagged underdeveloped web/mobile continuity. |
| Hyperdev / matsuoka.com (Jun 4, 2025) | Critical review | "Promising concept, premature execution" — slow response times, claimed-but-broken builds, awkward multi-window UX; positive on architecture and MCP bridge. |
| Tembo CLI tools comparison (2026) | Pro review | Frames Droid as the strongest delegation-style agent vs. Claude Code's interactive loop; recommends pairing the two. |
| byteiota — $150M raise coverage | Industry signal | Factory's funding and positioning emphasize model-agnostic, multi-agent coding with explicit role boundaries (code, review, docs, test, knowledge). |
Decision surface
What to use instead
| If you want... | Use |
|---|---|
| Multi-surface ticket delegation across repos with BYOK | Factory Droid |
| The cleanest interactive Anthropic pair-programmer | Claude Code |
| OpenAI-native local + cloud coding with first-party tooling | Codex CLI |
| GitHub-only issue-to-PR automation | Copilot coding agent |
| IDE-native background agents | Cursor background agents |
| Open-source self-hosted coding-agent platform you fully own | OpenHands |
| Managed autonomous engineering with Devin-style budget | Devin |
| Hackable terminal harness for harness-level customization | Pi |
Caveats
Where Droid can surprise you
- Droid is a closed runtime. Public docs describe enterprise hybrid and airgapped deployments, but there is no primary-source evidence the runtime itself is open source.
- BYOM Droid Computers route through `relay.factory.ai` by default, which is outbound-only but still a Factory cloud dependency unless you are on the airgapped enterprise path.
- Custom models via BYOK currently work in the CLI only, not on Factory's web or mobile clients.
- Background runs surface mistakes at PR review time, including occasional false-success claims on broken builds; pair Droid with strong CI gates and review.
- The web/mobile experience is real but reportedly not yet a clean continuation of local sessions, so don't assume seamless laptop-to-phone handoff.
- Enterprise pricing, concurrency limits for `droid exec`, and exact airgapped packaging are not published; treat them as diligence items, not given facts.
- Vendor-reported Terminal-Bench scores were submitted by Factory and should be replicated against your own task set before being treated as decision-grade evidence.
Cheat sheet
Quick reference
| Question | Answer |
|---|---|
| Install on macOS/Linux | `curl -fsSL https://app.factory.ai/cli | sh` |
| Install on Windows | `irm https://app.factory.ai/cli/windows | iex` |
| Install via npm | `npm install -g droid` |
| Run interactively | `cd <repo> && droid` |
| Run once in CI | `droid exec --auto low "add JSDoc to all public functions"` |
| Stream events to a pipe | `droid exec "complex task" --output-format stream-json | jq -r 'select(.type == "tool_call")'` |
| Add a project subagent | Create `<repo>/.factory/droids/<name>.md` with YAML frontmatter and a system prompt body. |
| Add a reusable skill | Create `<repo>/.factory/skills/<name>/SKILL.md` (or `skill.mdx`) with `name` and `description` frontmatter. |
| Bring your own key | Configure Anthropic, OpenAI, or an OpenAI-compatible gateway and switch with `/model`. Keys stay local. |
| Pricing entry point | Pro $20/mo. Step up to Plus $100/mo for Droid Computers and ~5x usage; Max $200/mo for ~10x and early features. |
How this was built
Sources, freshness, version
This guide is verified against Factory's product site, public documentation (CLI overview, droid exec, BYOK, Custom Droids, Skills, BYOM, enterprise network and deployment), the September 2025 Terminal-Bench announcement, and independent reviews from Every and Hyperdev/matsuoka.com on 2026-05-06. Pricing, model coverage, and benchmark rankings change frequently — re-check before adoption decisions.
Receipts
Claim table
| Where | Claim | Confidence | Source | Caveat |
|---|---|---|---|---|
| Bottom line | Factory positions Droid as software-development agents that work everywhere developers do — terminal, IDE, desktop, web, Slack, and integrated PM tools. | High | Factory.ai | Vendor positioning, not independent capability proof. |
| Snapshot | Droid installs via `curl -fsSL https://app.factory.ai/cli | sh`, `irm https://app.factory.ai/cli/windows | iex`, or `npm install -g droid`. | High | Factory.ai | Install URLs and npm package shape may change with major releases. |
| Autonomy tiers | `droid exec` defaults to read-only and gates mutation behind --auto low, --auto medium, --auto high, or the explicitly unsafe --skip-permissions-unsafe flag. | High | Factory.ai | Behavior of each tier is documented but evolves; verify against the current docs before automating high-risk pipelines. |
| Where Droid runs | Droid Exec supports text, json, and stream-json output and is intended for CI, scripts, and batch work. | High | Factory.ai | stream-json schema can evolve across releases. |
| Models and BYOK | BYOK supports Anthropic, OpenAI, and OpenAI-compatible providers including OpenRouter, Fireworks, Together AI, Ollama, vLLM, Baseten, DeepInfra, Google Gemini, Groq, and Hugging Face; only Anthropic and OpenAI are fully tested. | High | Factory.ai | Custom models via BYOK work in the CLI only, not on web or mobile clients. |
| Building blocks | Custom Droids are subagents defined in Markdown with YAML frontmatter, discovered from `.factory/droids/` (project) and `~/.factory/droids/` (personal). | High | Factory.ai | Frontmatter fields can be added or renamed in newer releases. |
| Building blocks | Skills live at `.factory/skills/<name>/` with `SKILL.md` or `skill.mdx`, and `.agent/skills/` is supported for compatibility. | High | Factory.ai | User-invocable behavior depends on the `user-invocable` and `disable-model-invocation` frontmatter flags. |
| Building blocks | BYOM lets users register Linux, macOS, or Windows machines as Droid Computers, with the daemon connecting outbound to relay.factory.ai when started with `--remote-access`. | High | Factory.ai | BYOM is not the same as the airgapped enterprise path; relay still depends on Factory infrastructure. |
| Enterprise deployment | Factory's enterprise docs describe cloud-managed, hybrid, and fully airgapped deployments, with the data plane (model traffic) staying in customer infrastructure across all three modes. | High | Factory.ai | Exact components per mode depend on enterprise contract scope. |
| Snapshot | Public Factory pricing lists Pro at $20/mo, Plus at $100/mo, Max at $200/mo, plus Teams (≤150 seats) and Enterprise (unlimited, on-prem, dedicated compute) on custom pricing. | High | Factory.ai | Plan structure and usage caps change; per-mission, per-hour, and BYOK quotas are not fully published. |
| Reviews and user signal | Factory reports a 58.75% Terminal-Bench score (state-of-the-art per Factory) with Opus 4.1 at 58.8%, GPT-5 at 52.5%, and Sonnet 4 at 50.5%. | Medium | Factory.ai | Vendor-submitted benchmark; replicate against your own task set before treating it as decision-grade. |
| Reviews and user signal | Every's Vibe Check reports the author canceled Claude Max and ChatGPT Max plans for Droid, citing model switching and stronger plan mode, while flagging an underdeveloped web/mobile experience. | Medium | Every — Vibe Check | Single-author opinion piece, not a controlled benchmark. |
| Reviews and user signal | Hyperdev/matsuoka.com calls Factory CodeDroid "promising concept, premature execution," citing slow response times, claimed-but-broken builds, and confusing multi-window UX. | Medium | Hyperdev / matsuoka.com | Posted June 2025; product has shipped many releases since, so re-test current behavior. |
| Reviews and user signal | Independent CLI tools comparisons frame Droid as the strongest delegation-style agent versus Claude Code's interactive loop and recommend pairing them. | Medium | Tembo | Comparison snapshots age quickly as agents iterate. |
| Reviews and user signal | Factory raised $150M in 2025 with a model-agnostic, multi-agent coding-platform pitch built around role-specialized droids. | Medium | byteiota | Funding coverage; treat as industry signal, not capability proof. |
Sources