Claude for Agencies

Claude for agencies means each client gets a private workspace that Claude reads from and writes back to, so audits, reports, and outreach run on that client's real context, and a monthly report is generated from a ledger instead of rebuilt by hand. It is the business-operating-system idea planned to a book of business: instead of re-explaining twenty accounts to a tool every week, you give the tool a place to remember each one, with a wall between them.

The reason this matters is a capacity number. Agencies say they want to move from about 35 accounts per strategist to 64, an 83% increase, while teams already spend more than a quarter of their time on manual campaign and reporting work (AgencyAnalytics, 2025). You cannot reach that ratio by asking people to work faster. Strategy is not the gap. The manual work around strategy is the gap, and that is the part a per-client workspace removes. For the underlying model this is built on, see Claude as a business operating system.

This guide explains what agency mode adds, how a workspace per client keeps one client from bleeding into another, how reports get generated instead of assembled, what the confidentiality firewall actually is and is not, and how to set the whole thing up in about thirty minutes with no code.

• Agencies want to go from about 35 to 64 accounts per strategist, an 83% jump (AgencyAnalytics, 2025).
• Teams already spend more than 25% of their time on manual campaign and reporting work (AgencyAnalytics, 2025).
• Agency mode adds two commands: /client (onboard, brief, prep, log, report, health, lifecycle, archive) and /team (sync, seats, onboarding, offboarding).
• Each client gets a private workspace with a confidentiality firewall, so client A never bleeds into client B.
• The lever is removing manual work, not working faster. Install runs about five minutes; the setup interview runs about thirty.

Why Is Account Capacity The Real Constraint?

Because the thing that caps a strategist is not how fast they think. It is the overhead stacked around the thinking: pulling reports for twenty clients, re-explaining each account to a tool that forgot it last week, rebuilding the same audit every month, hunting for last month's numbers before a call. That overhead is fixed cost per client, and it is what sets the ceiling on how many clients one person can hold without the work slipping.

The benchmark data names the gap directly. Agencies report wanting to run roughly 64 accounts per strategist against an actual figure near 35, and the same study finds teams spending more than a quarter of their hours on manual campaign and reporting work (AgencyAnalytics, 2025). Those two facts are the same fact from two angles. The accounts-per-strategist number is low because the manual-hours number is high. You move the first by cutting the second.

This is why the framing matters. "More accounts per strategist" sounds like a demand to work harder, and if you try to hit it that way the quality drops and people burn out. It is a capacity play, not a speed play. The job is to take the repeatable, manual parts of running an account and hand them to a system that remembers each client, so the strategist spends their hours on judgment instead of assembly.

What Does Agency Mode Add?

You turn on agency mode by running /setup and choosing Agency when it asks. That adds a Team/ area and a Clients/ area to the vault, and it adds two commands on top of the standard skill set.

The first is /client, the engagement manager. One command handles the whole arc of an account: onboard a new client or prospect, brief yourself on a client in one screen, prep for a meeting, log a meeting transcript so it compounds, draft a deliverable in the client's voice, generate a weekly or monthly report, run a health check across the book of business, move a client through the lifecycle, and archive an engagement with the lessons harvested on the way out.

The second is /team, the rollout manager. It sets up file sync between machines, creates seats with per-client access so each person sees only their accounts, generates the per-machine ignore files that enforce that planning, and onboards or offboards people cleanly. Everything else, the audits and the content and the lead work, is the same open skill set the rest of Conversion OS uses. Agency mode does not replace it; it wraps a client layer and a team layer around it.

How Do Per-Client Workspaces Work?

Each client gets its own folder under Clients/ with its own context: who they are, what they sell, their goals and baselines, their voice, the people on their side, the meeting history, the work in progress. When you run a skill inside a client workspace, it reads that client's context and nothing else. An audit for one client runs against that client's numbers. A draft for one client is written in that client's voice, not a house template and not another client's tone.

The onboarding flow builds the workspace by interview. /client new asks, one question at a time, who the client is and what you were hired to do, what success looks like in numbers, how they sound, who their customers are, who the key people are, and how the commercials work. It sets the status from reality, prospect if you are pitching, onboarding if they just signed, active if you are already serving, and it populates the goals file only with metrics the client actually stated. It never invents a baseline.

From then on the workspace compounds. Every meeting you log files to a dated note, routes decisions to a decisions log, turns actions into tracked tasks, and writes any metric the client mentioned into the goals ledger. Every deliverable lands in a dated drafts folder inside the client's work area. The point is that month two starts where month one ended, per client, without anyone re-keying the context.

How Do Monthly Reports Stop Eating The Month?

A report built from a ledger is generated, not assembled. Because every audit, result, and metric mention is written back to the client's workspace with a date, the report skill reads the ledger and produces the document: numbers first, narrative second, deltas measured against the real baseline that is already on file. /client report writes a weekly or monthly file straight into the client's reports folder, then offers a cover note for the client email.

That changes the economics of the most-hated task in the agency week. The manual version means opening five tools, screenshotting dashboards, pasting them into a deck, and writing the same "here is what we did" paragraph twenty times. The generated version reads what already happened, because the logging step captured it as it happened, and renders the report from the ledger. The strategist edits and sends instead of building from scratch.

It also makes the report honest, which matters more than the time saved. A report assembled from screenshots can quietly skip the metric that went the wrong way. A report generated from a ledger charts movement against a dated baseline, and where a goal has no fresh data it says "no new data" instead of guessing. You get a document that stands on its own and that you can defend on a call, because every number in it traces back to a line in the workspace.

One command can keep the workspaces current between reports. /autopilot sets up a scheduled agent that files meetings, chat, and email into each client workspace as they land, so the ledger is never stale when the report runs. It is optional. You can log by hand and the reports still generate. Autopilot just means the context maintains itself instead of waiting for you to catch up before month-end.

What Is The Confidentiality Firewall, Exactly?

The firewall is the rule that keeps client A out of client B. Every client workspace carries a plan contract: while Claude is working inside Clients/{client}/, it reads only that folder plus the shared library and company standards, it never reads a sibling client, it uses that client's own voice file, and it writes only inside that workspace apart from two narrow exceptions, a one-line daily activity entry and a task assignment. No other client's work may be referenced, quoted, or adapted into a deliverable. The vault's own audit lints for violations, cross-client references, escaped client material, credential patterns, and surfaces them above everything else.

It is worth being precise about what each layer is, because overselling this would be a mistake. There are three.

• The behavioral firewall, always on. Continuously enforced working discipline for the AI, with an independent audit. It is strong plan control for what the model reads and writes. It is not an access boundary for a human with the files open.
• Seat planning, for teams. Each person's access file lists their clients, and /team plan generates machine-local ignore entries so Claude on that machine never loads out-of-plan clients. On a folder-shared sync setup, the other clients' folders never even arrive on the machine. This is need-to-know by default and accidental-gap prevention. We do not sell it as access control, because it is not security against a determined human with filesystem access.
• Hard partition, for an NDA. When a contract requires real separation, that one client's folder becomes its own private repository, mounted into the vault only on the machines whose seats hold access. Enforcement is repository membership; revocation is removing it, in one step. The workspace spec is identical, so every skill still works.

The reason this is easier to reason about than most agency tool stacks is that the data is plain files you control, not rows in a shared database you cannot see into. You can open a workspace and see exactly what is in it, and exactly what the plan contract lets the AI touch. The default is the behavioral firewall plus seat planning; you escalate to a hard partition per client when a contract calls for it, not across the whole vault.

What Do /client And /team Look Like In Practice?

A normal week with a new account runs like this. You add the client with /client new and answer the interview. Before the kickoff you run /client prep, which gives you the open actions with their ages, the metric movement since the last touch, the decisions waiting on the client, and talking points in the client's own vocabulary. After the call you paste the transcript into /client log, which files the note, extracts the decisions, turns the actions into tracked tasks, writes any metric mentioned into the goals ledger, and updates the relationship notes. At month-end you run /client report and edit the generated document instead of building one.

Across the book, /client health all scans every active workspace, scores risk from real signals, and regenerates a pipeline view and a revenue view, flagging renewals inside 45 days even when an account looks green. When an engagement ends, /client archive runs the offboard checklist, harvests anonymized lessons into the shared playbooks for the user to approve, drafts a consent-flagged case-study candidate from the final ledger numbers, and moves the folder into a dark archive the AI can no longer see.

On the team side, /team sets up sync, recommends one option honestly rather than listing five, and creates a seat per person with an access file naming the clients they touch. It updates the roster and the relevant client team arrays, generates the planning file that controls what their machine loads, and hands them an onboarding checklist. Offboarding reverses it: reassign their open tasks, remove them from the client arrays, archive their folder, and revoke sync access first, because every later step assumes they can no longer pull files.

Try The Per-Client Idea In One Prompt

You do not need the full install to feel why a workspace per client beats a shared chat. Open Claude, paste the prompt below, and fill in one client's facts. It is the manual, one-shot version of a client brief: useful once, but with no memory, which is the whole reason the installed /client skill is worth setting up.

That prompt produces one brief, once. The installed skill reads a workspace you built during onboarding, writes the brief from a ledger that has been logging every meeting and metric, and remembers it next month so the following brief opens with the change. The prompt is the taste. The workspace is the system, and the system is what raises the accounts you can hold.

What Can It Do, And What Can It Not Do?

Honest limits matter more than a feature list, so here is the line. Agency mode is good at holding context per client, drafting in the right voice, scoring risk, and generating reports from a ledger. It is not a replacement for the strategist's judgment, and it is not a legal access-control system you can point at an auditor.

The right mental model is a tireless account coordinator who has read everything about each client, never confuses one with another while working, and never forgets a number. The strategist still sets direction and makes the calls. The system removes the manual assembly around them and shows the change, so the calls are better informed and the strategist has the hours to make more of them.

Manual Client Ops Or The Installed Skills?

The same agency can run either way. The difference is whether each client's context lives somewhere the tool reads and writes, or only in someone's head and a pile of dashboards. The table is the difference in practice.

How Do You Set It Up?

Conversion OS is a free, open set of these skills, and agency mode is part of it. Installing the plugin takes about five minutes and no coding. The one-time setup interview that builds your vault and your first client workspaces runs about thirty minutes.

1. Open Claude Code or Cowork, then Plugins, then Marketplace.
2. Add the Conversion OS repository and install it.
3. Run /setup and choose Agency when it asks. Claude interviews you and writes the vault, including the client and team areas, in one session of about thirty minutes.
4. Onboard one real client with /client new, then run /client brief on them to see a result land. After that, every skill uses the workspace you just built.

You do not have to wire up the whole book on day one. Most teams onboard one client, prove the loop on that account, brief, log a meeting, generate a report, and add the rest once they trust it. The point of week one is one client running on the system end to end, not a finished rollout.

What Are The Common Mistakes?

Treating a client workspace like a shared chat. If you paste everything into one conversation, you lose the firewall and the memory both. Onboard each client into its own workspace, or none of the separation or the compounding happens.

Skipping the logging step. Reports generate from the ledger, so a workspace that never gets meetings logged has nothing to generate from. Log the call, even briefly, or run autopilot, so month-end has something to read.

Overselling the firewall. The behavioral firewall and seat planning are strong attention control, not human access control. When a contract demands real separation, escalate that client to a hard partition. Do not promise an auditor what only the partition delivers.

Rolling out top-down to the whole team at once. Mandates stall. Run the vault solo for two or three weeks, add the most curious teammate next rather than the most senior, then expand once the value is visible.

Inventing baselines to fill a goals file. Populate goals only with numbers the client actually stated. A made-up baseline poisons every report that compares against it.

What Should You Expect In The First 30 Days?

Week one is setup and one client running, not a finished rollout. You run the setup interview in agency mode, onboard a single real account, and prove the loop: brief, log a meeting, generate a report. The win is small and concrete: that client now lives in a workspace with a baseline, so every future run has something to compare against.

Weeks two and three are habit, not a project. You log calls as they happen, add a second and third client when you trust the loop, and let the workspaces accumulate context. If the manual logging starts to drag, turn on autopilot so the context maintains itself. None of this is heavy. It is a few minutes around work you were already doing, except now the meetings and metrics are saved instead of scattered.

By the end of the first month you generate a report that opens with the change since the last one, across more accounts than you onboarded in week one, in less time per account than the manual version took. That is the moment the capacity math starts to show up: not because anyone worked faster, but because the manual overhead per account came down. If you do one thing, make it that first generated report off a real ledger.

Methodology

The capacity and time figures here come from the AgencyAnalytics 2025 marketing-agency benchmarks, linked above: the stated target of roughly 64 accounts per strategist against an actual figure near 35, and teams spending more than a quarter of their time on manual campaign and reporting work. Conversion System publishes no client outcome numbers, so none are claimed in this article. The /client and /team descriptions, the workspace model, and the three-layer confidentiality model reflect the open Conversion OS skills, which you can read in the public repository. The aim is to help an agency decide whether each client's work is attached to a real, separated context before trying to raise accounts per strategist.

Related: Claude as a business operating system · 12 Claude skills for small business · Claude for lead generation and scoring · Win clients with a free revenue audit