Skip to content

Agents for people who run more than one company.

Run every company's agents from one place.

Cohold holds the rules, decisions and numbers your agents work from, runs their routines on a worker you choose, and passes Slack and X actions through a gate with an allow list and limits that records requests, refusals included.

Preview of the panel with example data. Routines and runs are live, and the action gate checks the Slack step. Handing a task from one agent to another is in development.

You set an agent's autonomy level. Today the gate limits what leaves a run.

You choose each agent's autonomy level, and a new agent starts at Ask. Applying levels to what an agent does while it runs is in development. Today the limit on what leaves a run is the action gate: an allow list, hourly and per-run limits, and a record of requests, refusals included.

One memory, read by agents according to their scope.

A record has a scope: organization, company or area. A write policy decides how a new record takes effect: straight away, reviewed afterwards, or only after it is approved. What an agent writes is in force by default and waits in a review list. When a record changes, the records that depend on it are flagged for review.

  • LiveScopes and write policy
  • LiveReview flags on dependent records
  • LiveContext preview per scope
Records
  • RuleWeekly report goes out on Tuesday before 10:00
  • DecisionHarbor shop pauses paid ads in January
  • MeasureTarget cost per order for Lumen
  • RuleA missing value is shown as a dash, never as zero

Context preview

Computed with your own access. Opening it does not count as a load.

  • Organization: 3 records
  • Company Harbor: 5 records
  • Area Reports: 2 records

Example data. The companies and numbers are made up.

Agents from role templates. Routines on a clock.

Give an agent instructions, a model, tools, a memory scope and an autonomy level, then put it on a routine with a schedule and a time zone. Every run leaves a log you can read step by step. Role templates are installed by us today. Handing a task from one agent to another is in development.

  • LiveAgent editor and agent tree
  • LiveRoutines with a schedule and time zone
  • LiveRun log, step by step
  • In developmentMarketplace screen for role templates
  • In developmentHanding tasks between agents
Agent tree
  • Holding orchestratorOrganization
  • Reports agentHarbor
  • Ads agentLumen
Weekly routines
  • Mon08:30 · Daily budget check
  • Tue09:00 · Weekly report
  • Wed08:30 · Daily budget check
  • Thu
  • Fri16:00 · Memory upkeep

Example data. The companies and numbers are made up.

Set how much an agent may do on its own.

Each agent has an autonomy level: Draft, Ask, Do and report, or Silent. A new agent starts at Ask. You can also write a rule per action type. Both are saved today. Applying them to what an agent does while it runs is in development, and until then no level or rule changes what an agent can do.

  • In developmentFour autonomy levels per agent, saved; applying them is in development
  • In developmentAction rules, saved; applying them is in development
  • LiveApproval cards for memory proposals, verification and reading from Meta and BigQuery
  • LiveEdit before approving, on memory proposals only
  • In developmentApproval card before a Slack, X or Jira action
  1. 1Draft
  2. 2Ask
  3. 3Do and report
  4. 4Silent

Post the weekly summary to #harbor-reports

Planned card for a Slack post, not built yet

ApproveReject

Default action rules (saved; applying them is in development)

  • MoneySet to always ask
  • DeletionSet to always ask
  • Public postSet to ask, unless its check passed
  • Reading and reportsSet to run on its own

Example data. The companies and numbers are made up.

Actions leave through one door.

Slack, X and Jira actions go through the gate. The gate checks the request against an allow list of target types and destinations, applies hourly and per-run limits, and records requests, refusals included. There is no approval step for each post today: when sending is switched on for that routine and on the server, a Slack action goes out without anyone approving it. X posts are in dry run: the gate records the payload and calls nothing. Jira actions go through the gate too; whether they are sent for real or kept as a dry run depends on the server setting. The credentials used to send stay in the vault and never reach the run.

  • LiveGate with allow list, limits and request log
  • LiveSlack messages through the gate
  • Dry runX posts through the gate
Action · Destination · Result
  • Slack message#harbor-reportsAllowed
  • X post@harbor_shopDry run, nothing sent
  • Slack message#randomRefused: not on the list

Example data. The companies and numbers are made up.

Runs on a worker you choose.

A runner pulls its jobs from Cohold, so it needs no open inbound port. Run it on a server you own, or use a worker managed by Cohold, which an operator switches on for each organization. Cohold does not write a secret to its log or its record, and the value lands in the run's environment. Today the worker on your server can still reach the internet directly. Sending everything through the egress proxy is in development, and so is a screen that installs a runner for you. For now we issue the install code.

  • LiveManaged worker, switched on per organization by an operator
  • LiveRunner on your own server, install code issued by us
  • In developmentEgress through the proxy only
  • In developmentSelf-serve runner install screen
2 of 2 healthy
  • harbor-officeYour serverIdle1 min ago
  • managed-1Managed by CoholdRunningjust now

Example data. The companies and numbers are made up.

Your model accounts, or ours.

Connect your own API keys from Anthropic, OpenAI or MiniMax, or use Cohold credit. A request on your account uses your key from the vault and does not draw down your credit. You set a fallback chain: when a step hits a rate limit, a model provider error, an empty credit or a connection problem, the next step takes over. Each run shows what it cost, as an approximate figure. For your own accounts it is an estimate at list price. For runs paid with Cohold credit it rests on the supplier's cost and an exchange rate, both approximate. Accounts are added by us today. A settings screen is in development.

  • LiveOwn accounts and fallback chain
  • LiveCost per run, approximate
  • In developmentSettings screen for your own accounts
Fallback chain
  1. 1Your Anthropic accountFalls through on: rate limit, model provider error
  2. 2Cohold creditLast step

Run cost (estimate): $0.01

Example data. The companies and numbers are made up.

Free to start. Pay for seats, not for your own models.

Planned: Free at $0, Team at $20 per member per month billed yearly, Enterprise by conversation. Your own model accounts work on every plan.

  • Free

    $0

  • Team

    $20 / member / month, yearly

  • Enterprise

    Talk to us

Prices may change before launch. No one is charged during early access.

What's live, what's being built.

This list shows what is live, what is a dry run and what is in development.

LiveLive today9
  • One memory with scopes, a write policy and approvals
  • Routines, runs and step-by-step run logs
  • Approval cards for memory proposals, verification and reading from Meta and BigQuery
  • Agent editor and agent tree, with an autonomy level saved per agent
  • Action gate with allow list, limits and request log (Slack)
  • Model accounts and fallback chain, cost per run
  • Managed worker (switched on per organization by an operator) and runner on your own server
  • Secret vault, with read-only proxies for ad and warehouse data
  • Chat with the panel and the agents
Dry runDry run1
  • X posts through the gate
In developmentIn development7
  • Role template marketplace screen
  • Autonomy levels and action rules applied while a run acts
  • Approval cards before Slack, X and Jira actions
  • Task handoff between agents, task queue and orchestrator
  • Settings screen for your own model accounts
  • Self-serve runner install and egress through the proxy
  • Online payment and credit top-up (an operator loads credit today)

Updated 15 November 2026.

Questions

What is Cohold?
Cohold is a workspace for AI agents across several companies. It keeps one memory of rules, decisions and numbers, runs agent routines on a worker you choose, and sends Slack, X and Jira actions through a gate with an allow list, limits and a request log. Autonomy levels and approval before an action are in development.
Who is it for?
People who run more than one company or brand: agency owners, founders with several businesses, and the small teams around them.
Where does my data live?
Records, runs and the vault live in Cohold's own deployment, and secrets sit in an OpenBao vault. Every model request, memory content included, goes to the model provider. A request paid with Cohold credit passes through a model intermediary on the way. We have not published the hosting location yet.
Can I connect my own model account?
Yes, with API keys from Anthropic, OpenAI or MiniMax. Your account's own limits apply, and Cohold does not raise them. Accounts are added by us for now.
What is Cohold credit?
A prepaid USD balance for model usage. Planned price: OpenAI Codex models at 50% of the official list price, Gemini models at 70%. It will cover Gemini and OpenAI Codex models. Claude and MiniMax work with your own API key (decided; switching Claude off for credit is in development). It is shared by the organization, and each company can have a cap. Whether unused credit expires is not decided yet.
Can an agent act without my approval?
Today, yes. Slack, X and Jira actions go through the gate, which has an allow list, hourly and per-run limits and a request log, but no approval step for each post. When sending is switched on for that routine and on the server, a Slack action goes out without anyone approving it. Autonomy levels and approval before an action are in development. New agents start at Ask.
When does it open?
There is no date. We invite in small groups, by fit rather than by sign-up order.
Is there Turkish?
The panel and this site come in English and Turkish. We have not set up a support desk yet.

Early access.

Tell us what the first agent should do. We invite in small groups, by fit rather than by sign-up order.

Which model accounts do you have?

We use your email to confirm your address and to invite you to Cohold. Your other answers help us decide whom to invite first. Details: Privacy Policy.

Analytics cookies are set only if you accept them.