Turn AI tasks into
business operations.

Boring AI is the operating layer for AI work a business owns. A webhook, email, form, schedule, or app event starts an agent in a shared workspace; sensitive steps pause for an assigned teammate; and each run keeps the reasoning, tool calls, inputs, outputs, approvals, and cost together in one reviewable record.

If the work is personal, ad hoc, or yours to steer, use what your plan includes: Claude, and the personal agents now included in many AI plans, handle that well. Boring earns its place when the workflow belongs to the business. It starts from an event rather than someone's chat, lives in a shared workspace where every change is a versioned diff, acts only through the exact actions granted to it, routes sensitive steps to an assigned reviewer, resumes eligible failed runs from the last completed step, and leaves one record per run that other people can inspect. You are paying for operating and governing the workflow, not access to another chatbot.

Supervised automation runs on its own for routine steps and stops for an assigned person on the ones that matter, with the boundaries enforced by the runtime rather than written into a prompt. You describe the work in plain language instead of drawing a flowchart. The agent can call only the exact actions you granted, and new agents start with every connected action waiting for approval until you decide what to loosen. At each pause, the assigned reviewer can approve the step, modify its exact arguments, or reject it, and the decision stays attached to the run that needed it.

Use them when the work is one person's. Each gives an individual an agent with a cloud computer to work on: you message it, it keeps working on schedules, and it checks with you before risky steps. Grok Bot's Team Bots, in public beta as of October 2026, let a team share one bot, though each routine still runs for the person who set it up. Boring is for work a team owns and answers for. A known event starts the agent: a webhook, a form, a schedule, an app event, or an API call. It lives in a shared workspace with roles, and every change to its definition is a versioned diff you can roll back. Sensitive steps go to an assigned reviewer who can approve, modify, or reject them, with due dates and reminders. Eligible failed runs resume from the step that failed, and each run keeps one record the team can review, on the model and keys you choose.

Those are flow builders: you draw a fixed graph of triggers and actions, and the graph is what runs. Boring has no graph — you describe the outcome and the agent plans the steps on each run, so a renamed field or an unforeseen edge case doesn't send someone back to re-wire nodes. The trade-off is real in both directions: for simple app-to-app plumbing a flow builder is cheaper and faster to set up, and Zapier's app catalog is far larger than our 1,500+ connected tools. Boring is for the workflows with judgment in them.

New agents start with every connected action waiting for approval; owners can loosen that to writes only or auto-approve eligible actions they trust, such as reads, and an action Boring can't classify is treated as a write. When a step needs approval, the run pauses before it executes and the request waits in the shared Tasks inbox for its assigned reviewer; a tool approval leads with the exact proposed call and its secret-scrubbed arguments. That teammate — or a workspace owner or admin — can approve, modify, or reject it in the console, or approve or reject it from the native mobile companion, and the agent's owner or an admin can reassign it; an approving note returns to the agent as context before the run continues. Gates can carry due dates and reminders, and runs that remain paused in the background follow the configured reject-or-skip outcome after the deadline and grace period. Every decision and handoff stays attached to the run.

1,500+ connected apps covering the common operations stack — chat, email, CRM, helpdesk, project tracking, finance, docs — plus guarded first-party packs for selected APIs, any HTTP API that ships an OpenAPI spec once an admin reviews the imported operations, and remote MCP servers that authenticate with a token. Agents can also load fully rendered public web pages, search knowledge you've added, and run short JavaScript or Python calculations in a sandbox with no network or credential access.

New agents run on Claude Sonnet 5.5 by default, through Boring-managed models with no API key to set up. Boring-managed models also include GPT, Gemini, and Grok with no key, and lower-cost open-weight models are rolling out on the same managed route. Teams that prefer their own provider accounts can bring an Anthropic, OpenAI, Google, OpenRouter, or compatible-endpoint key instead. Set a workspace default or pin a model per agent, and backtest the change with the agent's Confidence checks before you switch. If a saved key stops working, the run stops and names the provider rather than quietly moving your traffic to Boring-managed models.

A schedule, an inbound webhook, an inbound email, a public form submission, an event in a connected SaaS app, the Chrome extension on the page you're looking at, the workspace API, an MCP client such as Claude Desktop or Cursor, or another agent. Whatever starts it, the same agent definition, granted actions, and reviewers apply, and the run is recorded in the workspace rather than in one person's chat history. When a run fails you can retry it from the step that failed rather than from the top — provided it got at least one step done; the console tells you when only a full re-run is possible.

No. Nothing is fine-tuned on your workspace and nothing is shared across agents or workspaces. Agents do learn in a narrower, visible sense: each run ends with a recap, items that repeat get de-escalated instead of re-flagged every day, and a thumbs-down becomes a standing lesson. Those lessons are memory entries you can read and download as JSON, each showing the run that taught it, and per-agent learning can be set to automatic, review-before-it-applies, or off.

Yes. Each run has a trace you can open in the console: the agent's reasoning, every tool call with its inputs and outputs, every approval and who gave it, and the cost of the run. Those tool inputs and outputs are pruned after 90 days, along with artifacts and replay payloads; the run's outcome, output, and step list stay. Reports and dashboards an agent builds are saved as self-contained pages, and each agent has one stable link showing its latest one, retained 90 days. Any files the run wrote — a CSV, a JSON extract, a chart — are listed on the run and download from there; those stay inside the workspace rather than behind a public link. Run exports (CSV or JSON) and workspace CSV exports cover summaries rather than the full step-by-step payload record, and run exports hold the most recent 100 runs. There's no immutability or attestation story yet — so this is a trace your team can review, not a compliance-grade audit archive.

We're in early access, so terms are set per design-partner engagement instead of published on a pricing page. There is no per-step meter. Ask us and we'll be specific about your workflow rather than pointing at a table we haven't earned yet.

Yes — white glove is the other half of the product. Our team maps your operations, builds the agents on the Boring platform, and runs them with you. The agents live in your workspace, written in plain language your team can read, edit, version, and roll back, so the engagement can end without the automation ending.