The minimum viable harness

Most Bosses we talk to think they need to buy a harness. They don't. They already have one. They have not assembled it on purpose.

The harness is the Boss's stack: the model, the runtime, the memory, the integrations, and the chain of command that holds it together. Some of those pieces cost real money. Most of them you already pay for. None of them are BossMode. We deliberately do not sell the harness, and the rest of this essay is about why that is the architecture, not the gap.

What is a harness?

A harness is the Boss's working environment for AI work. It is not one tool. It is the set of tools that, together, let an agent receive a prompt, execute against your business data, and ship a result you can sign off on.

If you have ever pasted a prompt into ChatGPT, copied the answer into a doc, then pasted that into a CRM by hand, you do not have a harness. You have a clipboard. The clipboard is the dominant AI workflow in 2026 and it is the reason most "AI rollouts" fail by month two: nothing persists, nothing chains, and the operator is the duct tape between every step.

A harness, by contrast, holds prompts, recipes, memory, and approvals as one operating system. The Boss stops being the duct tape. They become the one who tells the harness where to point.

You are the Boss. You tell the Co-pilot where to go.

What are the five components every harness needs?

Five pieces. If any one is missing, the harness leaks.

1. The model surface. Where prompts actually run. Claude Code, Cursor, ChatGPT, Codex, Claude.ai. This is the engine. The Boss picks the engine they trust based on what they are doing. A Boss writing copy all day is happy in Claude.ai. A Boss shipping product is faster in Claude Code or Codex. We take no position on which one. The right engine is the one the Boss opens without thinking.

2. The recipe runtime. Where Packs execute. For most Bosses starting out this is the same surface as the model: Claude Code runs the recipe, Claude Code calls the model, same window. Once volume grows, the runtime moves to a dedicated tool like n8n, Zapier, Make, or shell scripts on a tiny VPS. The runtime is what makes "Monday at 7:14 a.m., run the outbound recipe" a thing that actually happens, every Monday, whether you remember it or not.

3. The memory layer. Where context persists between runs. At minimum, a file system: a folder of markdown with the Boss's voice corpus, offer file, avatar file, and standing orders. Better, a vault. Notion, Linear, Obsidian, or a structured markdown repo synced across devices. The memory layer is the difference between an agent that knows you and an agent that meets you for the first time on every prompt.

4. The integration fabric. How the harness talks to the rest of the business. MCP servers for the things that have them. API keys for the things that do not. Webhooks for inbound events. OAuth for the SaaS the Boss already pays for: the calendar, the inbox, the CRM, the payments processor. This layer is the difference between an agent that drafts an invoice and an agent that ships the invoice and watches for the payment.

5. The Cockpit. Chain of command, approval queue, and audit. This is the one we sell. The Cockpit is what keeps the harness from going off-leash: which agent is allowed to ship without review, which agent has to wait for the operator's thumbs-up, who saw what, who signed off, who gets the receipts when the client asks "wait, when did this get sent?" without it, you have a harness that runs without supervision, which is exactly the configuration that produces the month-two abandonment story.

A harness without the Cockpit runs. It runs unsupervised. Most Bosses discover the cost of that the hard way, around week six.

What does the $25/mo starter stack look like?

This is the floor. If you cannot put together a stack at this price, the problem is not the harness, it is that you have not committed to running on AI.

• Model surface: Claude Code Pro at $20/mo. One operator. Opus access via the underlying Pro tier. Every recipe in this essay runs on Pro.
• Recipe runtime: Same window. Claude Code is a real shell; you can wire cron, file watchers, and triggers without leaving it. For anything heavier, run n8n's free Community Edition on the same machine. Unlimited executions, no licence.
• Memory layer: A folder. Literally. ~/bossmode/memory/ with five markdown files: voice.md, offer.md, avatar.md, standing-orders.md, decisions.md. Versioned by git, free.
• Integration fabric: Whatever MCP servers and API keys the Boss already owns. Calendar, Gmail, Stripe, the CRM. No new integration spend.
• Control plane: BossMode Free. One workspace, one device, local dashboard, your rules. Cloud sync is not in this tier; that is the only reason to leave it.

Total cash out: $20/mo plus sweat. The Lead Rescue is free on packs.bossmode.ing, so the recipes are free too. Most Bosses on the Free plan are running the harness for less than the cost of one lunch with a prospect.

What this stack will not do: keep state synced across multiple machines, support a second operator, or run the heavier weekly Packs without you babysitting them. Those are the upgrade triggers, and they are the reason for the next stack.

What does the $150/mo serious-Boss stack look like?

This is what most paying Bosses settle on within their first quarter. The stack assumes the Boss now wants the harness running while they sleep.

• Model surface: Pick one. Claude Code Max at $100/mo for the Opus-heavy Boss. Or Cursor Pro+ at $60/mo plus ChatGPT Plus at $20/mo for the Boss who still does a lot of writing in a separate window. Either configuration runs the same recipes; the choice is a temperament question, not a capability one.
• Recipe runtime: n8n Cloud Pro at €60/mo, roughly $65 once Stripe takes its cut. 10,000 executions a month, unlimited workflows, unlimited integrations, unlimited users. n8n is the cheapest serious runtime on the market by a wide margin: Zapier's Professional plan is $29.99/mo for 750 tasks, which sounds cheaper until you do the multiplication on a 10-step Zap.
• Memory layer: A real vault. Notion or Obsidian, depending on whether the Boss likes to read in a browser or in a local app. Cost: free to ~$10/mo. The vault holds the same files as the starter stack, plus an indexable archive of approved drafts, sales call transcripts, and the ledger of every commitment the agents have made on behalf of the business.
• Integration fabric: Same OAuth and API keys as before, now wired through n8n so the agents can read and write without the Boss opening a tab. MCP servers where they exist; webhooks where they don't.
• Control plane: BossMode Cockpit access through Fly AI. Three workspaces, unlimited devices, approval queue, audit export. Cloud sync turns on at this tier and stays on. The approval queue is the thing that makes the harness sleepable: you wake up, you see ten things waiting for a yes, you tap through them on the phone before coffee.

Total cash out: roughly $150 to $210/mo depending on which model surface you picked and whether you bought the n8n annual plan. That is less than one hour of a fractional COO at the low end of the band, the same fractional COO who, at $7K to $15K a month, would not be running your inbox at 6 a.m. on a Saturday.

This stack runs the standalone Packs cleanly: Outbound Engine (beta $197, was $497), Get-Paid Engine (Case Call-scoped), Lead Qualifier Engine (Case Call-scoped), and as many of the others as the Boss's bottlenecks call for. Recipes drop in. The runtime executes. The Cockpit catches everything that needs the Boss's eyes.

What does the $500/mo team stack look like?

Once the Boss has more than one human on the system, the stack changes shape. You are no longer optimising for one Boss's workflow; you are optimising for the team's shared memory and the boss's audit trail.

• Model surface: Claude Code Max at $200/mo, used across the team via shared seats. More often, two seats at $100/mo each so the Boss and the crew lead are not fighting over rate limits. Some teams add Cursor Pro+ for the developers and ChatGPT Plus for the writers; expect $200 to $400/mo total on model surface alone for a four-to-six-person operation.
• Recipe runtime: n8n Cloud Business at €800/mo (~$870), or a Make Teams plan if the Boss already has Make muscle memory. Business tier on n8n includes SSO, 40,000 executions, and the workflow-versioning that small teams stop being able to live without around month four. If $870 looks heavy, it is, until you compare it to the same volume on Zapier, where the same number of tasks will run you several thousand a month.
• Memory layer: A team-readable vault. Notion Team or a self-hosted Obsidian sync. The team can edit the same voice.md and offer.md the agents read; the agents see the edits within minutes. The vault is the team's institutional memory, not the boss's private notebook.
• Integration fabric: Same as serious-Boss, plus the sales tools the team actually uses. Pipedrive or HubSpot via API, Slack via MCP for chain-of-command messaging, Stripe webhooks into the ledger.
• Control plane: BossMode Studio at $149/mo for five team members and shared Packs, or BossMode Scale at $499/mo for unlimited team, advanced approvals, and SSO/audit when the boss is selling to clients who care about that. Cloud sync, shared memory, and the chain of command (Crew to Crew lead to Co-pilot) all live here.

Total cash out: $500 to $1,500/mo, depending on team size, runtime tier, and whether the boss has bought into Studio or Scale. That is less than the monthly retainer of a single AI consultant at the boutique end of the band ($150 to $300/hr translates to a $5K to $20K project for a build the team can now run themselves).

At this stack, the Trust Pack starts to make sense as one DFY engagement. Twelve component Packs, 48-plus standing orders, the entire OS pre-wired. Trust Pack runs $14,997+ DFY because the install requires the team's harness to actually exist, which at this stack it does.

What do we deliberately leave out?

We do not sell the harness. That is not a roadmap gap. It is the architecture.

We do not sell the model surface. The Boss picks Claude Code, Cursor, ChatGPT, Codex, or whatever comes out next year. The model market is moving fast and the Boss should not be locked into our choice of engine three years from now when something better exists.

We do not sell the runtime. n8n, Zapier, Make, shell scripts, raw cron: we ship the recipe, you pick the runtime. If we were the runtime, our recipes would only work inside our walls. They would also stop working the day we stopped maintaining them. Both of those are bad for the Boss.

We do not sell the memory layer. Your vault belongs to you. Your voice corpus, your offer file, your decisions log: all of it lives in your file system or your Notion or your Obsidian. The day BossMode disappears, your memory does not. The day a better vault tool appears, you can switch without losing a sentence.

The harness stays the Boss's. Always. The Cockpit is the one piece we sell, because the chain of command is the piece that turns five separate tools into one operating system. Everything else is yours.

This is also why the Pack catalog reads the way it does. A Pack is a recipe your harness executes, not a feature in our product. The Pack ships as a folder of standing orders, runs in your runtime, reads from your memory, hits your integrations, and reports up your chain of command. You own the install. You can read every recipe. You can fork it. You can fire us and the Packs keep running.

That last point is the one most SaaS vendors will not say out loud. We will say it because it is true and because Bosses figure it out anyway, and the ones who figure it out trust us more, not less.

The Harness Checklist (download)

We keep a one-page checklist of every component, every starter stack, and the upgrade triggers that move you from one tier to the next. It is the fastest way to look at your own setup and see which of the five layers is missing or held together with tape.

If your install is held together by clipboards and goodwill, take the Bottleneck Check first. Twelve questions, four minutes. It names the top three leaks in the business, not the harness, because the harness is the means, the leak is the bleed, and the Pack is the standing order that stops the bleed.