AI workers for multi-client agencies and operators

Demos are easy. Real client work needs more than a model.

Serve more clients with the team you already have.

Give every AI worker the right client context, APIs, accounts and permissions — without rebuilding access for every client.

Reuse what your agency learns. Keep every client's data and access isolated.

No copied keys. No .env files. Wrong-client access blocked.

Start 14-day trial Download for Mac

Create your trial. Download the Mac app. Run your first API proof locally.

Guided setup included · API keys stay local · Cancel anytime

outloop runtime access vault stays locked

One access setup Workspace approved Runtime allowed secret_exposed:false

Reuse approved API access across the runtimes your team already uses

One approved access layer for Cowork-style sandboxes, Claude Code, Codex, Hermes, and OpenClaw — without rebuilding setup for every platform.

One credential. The right workspace. Any approved agent runtime. secret_exposed:false

Independent tools. Names and logos belong to their respective owners; Outloop is not affiliated with or endorsed by these projects.

Agency capacity

More client work, without growing the team at the same rate.

Every new client used to mean another round of setup by hand — and the same question, every time.

Where is the API key?

Take me out of the damn API loop.

The same team running four client lanes at once First a person stands inside every client lane doing setup by hand and only one lane moves. Then the people step out, AI workers run all four lanes in parallel, and an approved lesson from a finished lane moves into a shared know-how tray that pre-loads the next lane. The team module stays the same size throughout. YOUR TEAM unchanged Client A method waiting · setup by hand Client B method waiting · setup by hand Client C method waiting · setup by hand Client D method waiting · setup by hand REUSABLE KNOW-HOW CLIENTS IN FLIGHT SAME TEAM 1 The same team running four client lanes at once The same scene as the desktop view: people step out of the client lanes, AI workers run all four in parallel, and an approved lesson moves into a shared know-how tray. YOUR TEAM unchanged Client A method waiting · setup Client B method waiting · setup Client C method waiting · setup Client D method waiting · setup REUSABLE KNOW-HOW CLIENTS IN FLIGHT 1

  • More delivery capacity

    AI workers take the repeatable execution.

  • Reusable agency know-how

    Approved lessons carry to the next client.

  • Less repeated setup

    Context and access stop being rebuilt.

  • Cleaner client boundaries

    Every client stays its own environment.

Context & learning

Share how the agency works. Never mix who it is working for.

Methods travel. Clients don't. A lesson only becomes shared knowledge after a person approves it.

Methods travel down to every client; client data never travels across An approved agency method flows down from a shared library into all three client workspaces. Client-owned data attempting to move sideways between two workspaces is stopped at a barrier with the denial code TENANT_MISMATCH. A lesson travelling back up to the library waits at a human review gate and only joins the library after a person approves it. AGENCY METHODS shared on purpose method process prompt lesson CLIENT A agency method reused accounts · files data · execution stays here CLIENT B agency method reused accounts · files data · execution stays here CLIENT C agency method reused accounts · files data · execution stays here TENANT_MISMATCH HUMAN REVIEW lesson proposed Methods travel down to every client; client data never travels across The same scene as the desktop view: an approved method flows down into both client workspaces, sideways movement between them is blocked, and a lesson travelling back up waits at a human review gate. AGENCY METHODS method process lesson CLIENT A method reused accounts data · files CLIENT B method reused accounts data · files TENANT_MISMATCH HUMAN REVIEW

Learning is a proposal, not an edit. No model training, no self-rewriting, and nothing learned for one client is applied to another.

How context and governed learning work

The AI worker

An AI worker needs more than a model.

An AI worker is an agent operating with the context, tools, permissions and working environment needed to complete real work. Seven things have to be true for one client task to run end to end.

  • Model and runtime
  • A computer to work on
  • Company and client context
  • Tools, APIs and files
  • Permissions
  • Reviewed learning
  • Audit

The first two are yours to choose. Outloop is not the model, not the agent runtime, and not the computer. It is the layer that keeps the other five true: which client an AI worker is working for, what it may use, and what happens to what it learns.

How Outloop runs one client task end to end An AI agent, a client workspace and approved client context feed into the Outloop workstation. Inside it, a request passes seven checkpoints — context, workspace, account, access, files, audit and learning — and then reaches an approved destination system such as Google Ads, Meta, Drive or GA4. A second request that names the wrong client is stopped at the account checkpoint and never reaches an external system. AI Agent Claude Code · Cowork Client Workspace One client per workspace Approved Context Brand rules · assets APPROVED WORKSPACE ACTION COMPLETED WRONG CLIENT ATTEMPT RESOURCE_ID_NOT_ALLOWED OUTLOOP The control layer 01 Context loaded blocked 02 Workspace identified blocked 03 Account bound blocked 04 Access approved blocked 05 Files routed blocked 06 Audit recorded blocked 07 Learning captured blocked Google Ads Meta Drive GA4 Business Systems How Outloop runs one client task end to end A request from an AI agent enters the Outloop workstation, passes seven checkpoints — context, workspace, account, access, files, audit and learning — and reaches an approved destination such as Google Ads, Meta or Drive. A request naming the wrong client is stopped at the account checkpoint and never reaches an external system. AI Agent in one client workspace OUTLOOP The control layer 01 Context loaded blocked 02 Workspace identified blocked 03 Account bound blocked 04 Access approved blocked 05 Files routed blocked 06 Audit recorded blocked 07 Learning captured blocked approved blocked Google Ads Meta Drive + Business Systems APPROVED WORKSPACE ACTION COMPLETED WRONG CLIENT ATTEMPT RESOURCE_ID_NOT_ALLOWED

Illustrative workflow preview — the seven checkpoints and the wrong-client denial are the model Outloop enforces, not a recording of a customer account.

  1. Context — approved client knowledge, retrieved before work starts.
  2. Workspace — the agent knows which client it is working for.
  3. Account — bound to the one account or resource it may touch.
  4. Access — uses the credential without ever seeing it.
  5. Files — moves real client files and media through approved paths.
  6. Audit — every attempt written to a redacted local log.
  7. Learning — a reviewed lesson improves the next run.
  8. A request that names the wrong client is denied at the Account checkpoint with RESOURCE_ID_NOT_ALLOWED, before any external system is called.

Stages 01–02

Start with the right context

Before it generates anything, the agent retrieves the approved knowledge for that client — brand rules, the assets it may use, what it must not claim. Work starts informed instead of starting blank.

Stages 03–05

Act through approved access

The workspace is bound to one client account. The agent requests an action, policy checks it, and the credential is used host-side. A request naming another client is denied before any backend call.

Stages 06–07

Close the loop

Allowed and denied attempts are both written to a redacted local audit. A lesson from the run becomes a Context revision only after a human approves it — and it can be rolled back.

Nothing here runs on its own judgement. The agent proposes, policy decides, and a human approves anything sensitive — including every change to what the agent has learned.

The access problem

Where real client work breaks.

Agents are easy in demos. Real client work breaks on access — the same five ways. Each line links to how Outloop handles it.

outloop access diagnostic5/5 handled

Right accessRight clientsecret_exposed:false

How it works

How you reuse API access in 3 steps

Add it once. Approve the workspace. Let the agent use it safely.

Outloop “Add an API key” panel: a “No terminal needed” badge, a service picker set to Google Ads, and a Workspace-dedicated access selector.
00

Add API access once

Choose a service, select the workspaces that should get access, and store the credential locally on the Mac.

Keys stay local
Outloop workspace approval: the outloop-website workspace selected to receive access, with a suggested key name and an empty “Paste the API key” field.
00

Approve the right workspace

Grant access only to the client workspace that should use it. Each workspace stays isolated.

Wrong-client access blocked
Outloop agent-projects panel: the Claude / Cowork runtime expanded to show per-project status (Needs action, Ready, Need to connect), above the Claude Code, OpenClaw, and Hermes Agent runtimes, with an “Agent keeps working — secret_exposed:false” proof badge.
00

Let agents use approved access

Connect agent projects, then let approved agents request access through Outloop without seeing the raw key.

Agent keeps working secret_exposed:false

Keys stay local Workspaces stay scoped Agents request access, not keys

The loop, before and after Outloop.

Before Outloop
  • Agent asks for a key.
  • A human pastes it in.
  • The key ends up in chat, .env, logs, or the wrong client folder.
  • Every new workspace repeats the same setup loop.
With Outloop
  • Agent requests approved access.
  • Outloop checks workspace policy.
  • The local broker performs the call.
  • Only a redacted result returns — the agent keeps working.

That's the API loop. Outloop removes it.

Start removing it

Agency workflow proof

Built from real agency API workflows.

Outloop was built while running real client-agent workflows across ads, CRM, data, file, reporting, and automation APIs.

The lesson was simple: agencies don't need another place to paste keys. They need one approved access layer that lets agents work across client workspaces safely.

Explore agency API workflows
Google Ads Campaign checks
Meta Ads Account reporting
Merchant Center Product feed review
Airtable CRM & ops data
Google Drive Client asset folders
Gmail Inbox workflows
Apify Data collection
Firecrawl Web research

Example services shown for workflow context. Logos and names are trademarks of their respective owners; no official integration or endorsement is implied.

Why we're building this

AI agents are becoming real workers. Real workers need a controlled workplace.

When a company hires a new employee, it gives that employee a computer. We think the AI worker is the new employee — and it also needs one: files, instructions, approved tools, and access to the systems the work actually lives in.

That is the part almost nobody is building. Models keep getting better at deciding what to do. Someone still has to decide what an agent is allowed to do, for which client, with which account.

Read the AI Agency Manifesto The Outloop assistant working at a desk

Keep your vault. Control runtime access.

1Password
macOS Keychain
Infisical
Doppler

Outloop works above Keychain, 1Password, Infisical, Doppler, and other secure backends. It does not replace your vault. It controls which workspace and runtime can use approved access.

  • No API keys uploaded to cloud.
  • No raw key returned to the agent.
  • No .env files required.
  • Wrong-client access is blocked before credential use.
Pricing

Pricing for multi-client agent operations.

Start with the workspace pack that fits your agency today. Upgrade when Outloop becomes part of how your team runs client AI workflows.

Pro

$79 /month

5 client workspaces

$15.80 per client workspace at full plan capacity

For solo operators and small dev shops running real client agent workflows.

  • ✓1 activated Mac
  • ✓5 client workspaces
  • ✓Up to 10 connected services
  • ✓30-day local audit history
  • ✓Email support
Start 14-day trial

Self-serve setup. No sales call required.

Secrets stay local on your Mac.

Recommended

Agency Growth

$499 /month

50 client workspaces

$9.98 per client workspace at full plan capacity

For established agencies running agents across many client accounts and workflows.

Everything in Agency Starter, plus —

  • ✓Up to 3 activated Macs
  • ✓50 client workspaces
  • ✓Up to 75 connected services
  • ✓90-day local audit history
  • ✓Priority support

Add-ons:

+10 workspaces for $60/month, maximum 3

+1 activated Mac for $29/month, maximum 2

Start 14-day trial

Need help planning your rollout?

Request a Workflow Review

Custom

101+ client workspaces

Tailored capacity for agencies operating at larger scale.

  • ✓Custom workspace capacity
  • ✓Additional activated Macs
  • ✓Custom connected-service limits
  • ✓Longer audit history
  • ✓Security review and DPA support
  • ✓PO and invoice purchasing

Guided implementation is scoped and quoted separately.

Contact Sales
See all plans →

Agency Starter ($249/mo, 20 workspaces) and Agency Scale ($899/mo, 100 workspaces) are on the pricing page.

Prefer annual? Billed yearly saves 18% — see all plans.

Secrets always stay local. Plans are based on activated Macs, client workspaces, and connected services. Outloop cloud never receives raw API keys, tokens, .env contents, Authorization headers, or secret-bearing files.

15-minute AI Workflow Review · Guided onboarding

Send us one real agent workflow. We'll show you where access will break.

A 15-minute AI Workflow Review of one workflow you actually run. We’ll review one real workflow from your agency, identify the tools, accounts, files and permissions it requires, and show you how it could become an AI worker workflow.

Ready to start right away? Start your 14-day guided trial → It begins on Outloop Cloud (account & license only); your API keys stay local on your Mac.

✓ We never ask you to paste secrets into this form.

In the review, we map:

✓ API setup & OAuth sessions
✓ MCP connections across clients
✓ Client separation & wrong-client risk
✓ Drive, folder & file access
✓ Browser-automation fragility
✓ Audit & proof needs

Built for agent workspaces like:

Claude Cowork logo OpenClaw logo Hermes logo Claude Code logo Codex logo

Independent tools. Names and logos belong to their respective owners; no official integration is claimed.

Request your 15-minute AI Workflow Review

Guided onboarding included · We reply by email · No card needed to talk to us

Or start the 14-day guided trial now

Guided onboarding · Secrets stay local · We never ask you to paste secrets into this form.

Do not paste API keys, passwords, Bearer tokens, Authorization headers, .env contents, or other secrets into this form.

By submitting, you agree that we may contact you about your Outloop guided trial and onboarding. Terms · Privacy · Refund Policy

Frequently Asked Questions

Got questions? We have answers.

Ready to get out of the API loop?

Serve more clients with AI workers — and reuse what works.

Put AI workers to work across real client operations — and turn approved learning into reusable agency know-how.

Start 14-day trial Request a Workflow Review

For agencies and operators managing 5 to 100 client workspaces.