# Factory > 🏭 Coding agent CLI engineered to make any LLM — local or cloud, frontier or 7B — actually useful. 16 providers, multi-tab sessions, automatic key+model failover, plan-mode review, MCP, and recovery loops for models that misbehave. - **Type:** MCP server - **Install:** `agentstack add mcp-vilaca-factory` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [vilaca](https://agentstack.voostack.com/s/vilaca) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** Apache-2.0 - **Upstream author:** [vilaca](https://github.com/vilaca) - **Source:** https://github.com/vilaca/factory - **Website:** https://vilaca.github.io/factory/ ## Install ```sh agentstack add mcp-vilaca-factory ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # 🏭 factory [](https://github.com/vilaca/factory/actions/workflows/lint.yml) [](https://github.com/vilaca/factory/actions/workflows/typecheck.yml) [](https://github.com/vilaca/factory/actions/workflows/test-unit.yml) [](https://github.com/vilaca/factory/actions/workflows/test-e2e-mocks.yml) [](https://github.com/vilaca/factory/actions/workflows/test-e2e-no-mocks.yml) [](https://github.com/vilaca/factory/actions/workflows/test-e2e-headless.yml) [](https://github.com/vilaca/factory/actions/workflows/test-e2e-pty.yml) [](https://codecov.io/gh/vilaca/factory) **A coding agent CLI engineered to make any LLM — local or cloud, frontier or 7B — actually useful.** ## Why factory - **Bring your own model.** Pick what fits your privacy, cost, and latency. 16 providers on equal footing — local-first ([Ollama](https://ollama.com), [llama.cpp](https://github.com/ggml-org/llama.cpp)) and cloud ([Anthropic Claude](https://www.anthropic.com), [Cerebras](https://cloud.cerebras.ai/), [Cloudflare Workers AI](https://developers.cloudflare.com/workers-ai/), [Codestral](https://mistral.ai/news/codestral), [Cohere](https://cohere.com/), [GitHub Copilot](https://github.com/features/copilot), [Google AI Studio](https://aistudio.google.com), [Groq](https://console.groq.com/), [HuggingFace](https://huggingface.co), [Mistral](https://mistral.ai), [OpenAI](https://platform.openai.com), [OpenCode Zen](https://opencode.ai/docs/zen/), [OpenRouter](https://openrouter.ai), [Vercel AI Gateway](https://vercel.com/docs/ai-gateway)). - Supports _Anthropic Claude-Code_ [skills](./docs/skills.md) format. - **Multi-tab sessions.** Each tab is an independent agent with its own conversation, working directory, provider, and model. Run a frontier LLM on a refactor in one tab while a local LLM explores tests in another. Switch with `Ctrl+N`/`Ctrl+P` or jump directly with `F1`–`F12`. - **Two-tier rotation.** When a key hits a rate limit or auth failure, factory swaps to the next saved key for the same model; when keys for a model are exhausted, it walks a configurable chain of `:` fallbacks — frontier → fast → free, automatic. - **Picker with key management.** Startup and in-session pickers support provider → key → model selection, key validation, and saved-key reuse for providers that support multiple API keys. - **Built for models that don't behave.** Text-tool fallback recovers tool calls from prose; an LLM corrector retries malformed calls; an imitation guard strips fabricated tool-result blocks; Bash dedup nudges the model out of spinning loops. - **Plan mode.** Read-only tools execute freely; writes are queued for review. Approve, cancel, or refine before anything touches disk. - **Cache-aware.** Surfaces prompt-cache hit rate per turn and per key for providers that report it. `/stats` reports session totals, per-turn hit-rate sparkline, compaction events, and the largest tool results. - **Human or headless.** Interactive TUI on a TTY; in scripts and CI it auto-detects no-TTY, reads stdin as a prompt, runs one turn, and streams the result to stdout — same agent, no UI. See [docs/headless.md](./docs/headless.md). ## Requirements - **Node.js >= 22** - At least one LLM provider — local ([Ollama](https://ollama.com), [llama.cpp](https://github.com/ggml-org/llama.cpp)) or cloud (see [docs/providers.md](./docs/providers.md) for credentials per provider). ## Quick start ```bash git clone https://github.com/vilaca/factory.git cd factory npm install && npm run build && npm link factory ``` That's it. `factory` opens a picker for provider, key, and model the first time (for simple-key providers, it can validate and save keys during the flow). Subsequent runs jump straight into the prompt with the last provider/model you used; pass `--pick` (or use `/model` / `Ctrl+K` mid-session) to choose a different one. > **`npm link` permission errors?** It writes a symlink into your npm global prefix; if that's a system path it needs sudo. Either set a user-writable prefix once (`npm config set prefix "$HOME/.npm-global"` and add `$HOME/.npm-global/bin` to your `PATH`), skip linking and run `npx factory` from the repo, or invoke directly with `node /path/to/factory/dist/index.js`. > **npm install coming soon.** A `factory-code` package will land on npm once the first tagged release ships; until then, build from source. ## Documentation - [docs/providers.md](./docs/providers.md) — full provider matrix, env vars, auth specifics - [docs/configuration.md](./docs/configuration.md) — CLI flags, env vars, config file format, experimental flags - [docs/slash-commands.md](./docs/slash-commands.md) — every `/command` and what it does - [docs/hotkeys.md](./docs/hotkeys.md) — keybindings - [docs/headless.md](./docs/headless.md) — non-TTY mode, exit codes, CI patterns - [docs/skills.md](./docs/skills.md) — reusable instruction sets, scopes, frontmatter reference, invocation - [docs/web-fetch.md](./docs/web-fetch.md) — WebFetch tool, bounds, per-domain whitelist - [docs/troubleshooting.md](./docs/troubleshooting.md) — common issues and fixes - [docs/observability.md](./docs/observability.md) — session-log JSONL schema - [docs/security.md](./docs/security.md) — built-in path jail, bash deny list, env scrubbing - [ARCHITECTURE.md](./ARCHITECTURE.md) — module map, data flow, design overview - [CONTRIBUTING.md](./CONTRIBUTING.md) — dev setup, conventions, adding a provider ## Quick reference ```bash factory --version # print version factory --help # full flag list factory # interactive picker factory -p anthropic -m claude-sonnet-4-6 # one-shot factory --plan # plan mode factory --pick # force startup picker factory --turn-timeout 120 # auto-abort after 120s factory --debug # debug logs to stderr echo "explain src/index.ts" | factory # headless / non-TTY ``` Once running, type `/help` to see the full slash-command list. ## Security Built-in path jail, bash deny list, env scrubbing for subprocess execution, and plan mode for untrusted models. See [docs/security.md](./docs/security.md) for what's enforced and how to extend it, and [SECURITY.md](./SECURITY.md) for the disclosure process. ## License Apache License 2.0 — see [LICENSE](LICENSE). ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [vilaca](https://github.com/vilaca) - **Source:** [vilaca/factory](https://github.com/vilaca/factory) - **License:** Apache-2.0 - **Homepage:** https://vilaca.github.io/factory/ Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** yes - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-vilaca-factory - Seller: https://agentstack.voostack.com/s/vilaca - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.