You bill $180/hr. Your stack costs you $400/mo and an hour of admin a day before you open a proposal.

Salesforce for the contact record. HubSpot for the follow-up sequence. Asana for project delivery. n8n to glue the tokens together. A wiki nobody opens after the second client. Each tool solved a real problem when you added it — and now the combination taxes the shop before any client outcome exists.

For a 2–25 person US consulting firm, that is not a technology problem. That is a structure problem. The tools give you data. None of them tell work what to do next, who approved the start, or whether an agent is allowed to act. You fill those gaps with Slack threads, personal judgment, and a weekly retro that always starts with "I thought we were done with that."

Cockpit is business structure in a box. Not a replacement for the AI you already run — Cursor, Claude, Grok do the work. Cockpit is the rails: desks with a promote-to-start gate, one approvals inbox for every high-stakes decision, and agent runtimes bound to a desk so nothing spins up unsupervised. @Supabase is the headless DB underneath — agents, CRM, and project management in one schema. @vercel is the UI layer. Cockpit is the control plane you run on top of that stack, not a replacement for either.

The stack tax in practice

Every tool in the pile was justified once. The problem is that they don't compose.

• Salesforce / HubSpot stores the contact. It does not know what the agent is allowed to do with that contact today, or whether the follow-up was approved.
• Asana tracks the task. It does not know whether a Cursor agent already started the work three hours ago from a Slack thread.
• n8n passes tokens between systems. Every workflow node holds a secret. That secret is a single engineer's responsibility to rotate, and it breaks silently.
• The wiki holds the knowledge. Nobody reads it when the agent needs context mid-session.

The tax is not any one tool. It is the re-key between them: the Asana card that duplicates the Salesforce record, the n8n webhook that references the Slack thread that references the email, the agent that starts because someone pasted the task into a chat instead of promoting it from a desk.

A 12-person boutique is not a startup with a budget for RevOps. The principal is often the CRM admin, the delivery lead, and the partner on the account. The stack should spend attention for them, not demand it.

What Cockpit actually ships

Here is what is live today — shipped, fair to name, no coming-soon.

Desks and projects. Marketing, meetings, ops, SDR — each desk has routines and a human promote/close loop. Work is not a shared chat scrollback. A task sits in backlog until a human promotes it to todo. A human closes it when done. Agents execute inside that lease; they do not invent a start. That is the pattern we wrote up in the Promote-to-Start how-to — the same gate works for client delivery desks.

Approvals inbox. Every high-stakes decision lands in one place: merge two contacts, start work, close work, approve an outbound draft, resolve a scorecard. Agents and ingest propose. You swipe in the inbox. The system of record updates after the decision, not before. For a consulting principal who is also the CRM admin, that inbox is the difference between a 30-second approve and a 30-minute archaeology dig when a contact went sideways. See the entity-match flow in the entity match how-to post.

Agent runtimes bound to a desk. Cursor Automation, Claude, and Grok each attach to a desk and run on its schedule. The agent does not invent its own cron. The desk owns the cadence. If you have ever untangled a Cursor session that ran unsupervised while the team was on client calls, this is the fix — not a new rule in AGENTS.md, but a hard runtime boundary.

Integrations + Vault. Google, X, and your other connections live on one account page. Connect once. Tokens sit in Vault, not in a .env file on a laptop or a n8n node that breaks on the next hire. The secret rotation problem becomes an account page problem.

Marketing and SDR desks. The same engine that writes this site and schedules X runs from a Marketing desk. SDR researches, qualifies, and drafts replies — outbound stays draft-and-hold until you approve. Both are live. The point for consulting firms is that these are not separate products: the same desk/promote/approve pattern handles client delivery, marketing, and outbound from one control plane.

What is not shipped: agent execution logs tied to live sessions (labeled coming soon on the product page), streaming agent UI inside Cockpit, Claude CCMA as a third runtime, client handoff embeds. Those are on the roadmap. This post stays inside what you can run today.

Why this is not another dashboard

Every AI dashboard claim sounds the same: "one place for all your agents." The difference here is the gate.

Cockpit does not show you agent activity after it happens. It controls what can start. Promote-to-start is the start signal. Approve-before-start is the contract that the agent respects. Draft-and-hold means the outbound email does not go without a swipe. Those are not dashboard features — they are the operating model.

[COCKPIT_WEDGE: No other system lets you fully automate the work and sell the agentic outcomes.] The flywheel is that marketing learns from work already completed in Cockpit, and you sell those completed outcomes. For a consulting firm, that means the same system that runs your internal operations becomes the proof that your implementation methodology works for a client — because you built it on your own shop first.

The stack it replaces (or shrinks)

You do not have to cut everything on day one. But the direction of travel is clear:

• HubSpot / Salesforce — CRM, pipeline, and contacts move to Supabase + Cockpit. Not another tab farm; people, companies, and the record of who decided what.
• Asana — Desks and promote/close replace the board. Work is not a card you update after the chat.
• n8n — Integrations tab + Vault. Connect once; tokens are not a laptop env file or a workflow node secret.
• Knowledge management — Git + hybrid search + multi-repo check-in. Company knowledge stays on the work item, not a wiki nobody opens before a proposal.
• Marketing automation — The marketing engine runs from a desk. It learns from completed Cockpit work, then sells those outcomes.

The run cost for the underlying stack: Vercel, Supabase, Google Workspace, Claude Teams, Cursor Ultra, Grok bots — roughly $300–500/mo for the agents and the host. That is not Cockpit's invoice; that is the stack you are running whether or not you add Cockpit. Cockpit is $0 until you sell a completed outcome to a customer.

Who this is for

Small US consulting firms: 2–25 people, principals and managing partners who sell hours and retainers. Shops that are buying the AI and learning it on their own work first — so that when a client asks "can you implement this for us?" the answer is "we already run it."

The implement-for-others step is step two in the journey. Step one is structure in your own shop: desk, gate, approve, close. That is what Cockpit gives you before the first client outcome.

See the product at cloudbeast.io/cockpit. Book a discovery call at cloudbeast.io/schedule if you want to map desks to your current tools and talk about what the first desk looks like for your firm.