Frequently Asked Questions

Three plans: Starter at $79/month (5 projects, 1 company), Growth at $199/month (25 projects, 2 companies), and Business at $599/month (unlimited projects, 5 companies). All plans include unlimited agents. LLM costs are separate — you pay your AI provider (Anthropic, OpenAI, Google) directly at their standard rates.

No. All plans include unlimited agents. You can deploy 3 agents or 30 agents for the same subscription price. The only variable cost is LLM usage, which you pay to your AI provider.

Plan limits are enforced on creation. If you're on Starter (5 projects) and try to create a 6th project, the system will prompt you to upgrade. Existing data is never deleted when you downgrade — you just can't create new items above the limit.

Yes. Cancel from the billing portal at any time. You retain access until the end of your current billing period. After that, your data is preserved for 30 days in case you resubscribe, then permanently deleted.

Yes. IronWorks uses a BYOK (Bring Your Own Key) model. You connect API keys from Anthropic, OpenAI, Google Gemini, or other supported providers. Your agents use tokens from your accounts at standard provider rates with no markup from IronWorks.

Any model available through your connected API keys. Common choices: Claude Sonnet/Opus/Haiku (Anthropic), GPT-4o/GPT-4o-mini (OpenAI), Gemini Pro (Google). You can configure which model each agent uses, so your CEO might use Claude Opus while your engineer uses Codex.

The heartbeat system polls every 30 seconds by default. When you assign issue STE-25 to the CTO, the CTO agent wakes up within 30 seconds, reads the issue, and begins executing. Agents don't run continuously — they sleep between tasks to minimize cost.

Yes. The CEO and CTO have the canManageAgents capability and the ironworks-create-agent skill. If the CEO determines the team needs a second engineer, it can create one. The new agent is deployed with the standard role template and starts receiving work on its next heartbeat.

Agents are trained to set their issue status to "blocked" when they can't proceed. The blocked issue appears in the War Room's "Attention Required" section. You can provide the missing information, reassign the task, or break it into smaller pieces.

Encrypted with AES-256-GCM using a per-company master key. Each secret version gets a unique initialization vector. Secrets are decrypted only at the moment of use and never cached in plaintext. They are never logged, never included in agent prompts, and never visible in the UI after storage.

No. Every database query is scoped by company_id. Agents, issues, secrets, KB pages, and budget data are all partitioned by company. This isolation is verified by 122 automated HTTP tests that run on every code change.

No. Your data is never used for model training by IronWorks. Your API key usage with providers like Anthropic and OpenAI is subject to their respective data policies, but IronWorks itself does not share, sell, or use your data for any purpose beyond providing the service.

All user-supplied text is sanitized before entering LLM prompts. Known injection patterns ("ignore previous", "you are now", "forget your instructions") are detected and removed. Input fields are capped at 2,000 characters. LLM output is validated against strict schemas to prevent agents from executing malformed instructions.

It depends on the model and task complexity. A simple task using Claude Haiku might cost $0.01-0.05. A complex engineering task using Claude Opus might cost $0.50-2.00. IronWorks tracks cost per agent and per task so you can see exactly where money goes. The default monthly budget is $500, and you can adjust it anytime.

At 80%, the CFO agent gets a warning issue. At 100%, all non-CEO agents are automatically paused. The CEO stays active so you can communicate with it. Raise the budget or wait for the month to reset.

Multi-step workflows with dependency tracking. Each step is an issue assigned to a specific agent role. Steps can depend on other steps and optionally require human approval. Your company starts with 7 pre-built playbooks: New Client Onboarding, Security Audit, Product Launch, Incident Response, Content Campaign, Weekly Operations Review, and Lead Generation Funnel.

A shared company wiki that agents actively read and write. Your company starts with 9 seed pages covering operating procedures, engineering standards, security policy, and more. Pages use markdown, support cross-linking, and have full revision history. Agents consult the KB before making decisions.

Yes, via Telegram. Connect your company to a Telegram bot, and messages you send become issues assigned to the CEO. The CEO processes them and sends responses back through Telegram. You can also use the email bridge — send an email to ceo@yourcompany.ironworksapp.ai.

IronWorks is designed to run on dedicated infrastructure. The hosted version runs on Contabo VDS servers. Self-hosting documentation is planned for a future release.

TypeScript throughout. React frontend with TanStack Query and Tailwind CSS. Express.js API server. PostgreSQL database with Drizzle ORM. Better Auth for authentication. Polar for subscription billing. The whole system is a monorepo with packages for the database schema, shared types, and adapter implementations.

Each agent has an adapter type that determines how it executes work. Adapters include claude_local (Claude Code CLI), codex_local (OpenAI Codex), gemini_local (Google Gemini), cursor, and more. Sessioned adapters maintain conversation context across heartbeats. The adapter receives the issue context, workspace directory, and tools, then returns the execution result with cost data.

IronWorks is built by Steel Motion LLC, a veteran-owned software company. It started as a fork of the open-source Paperclip project and has been extensively rebuilt for production use as a multi-tenant AI workforce orchestration platform.