FIG. 1AI agents for every team
Run your company with agents. Stay in charge.
OrchStack gives each team AI agents that do routine work inside the tools you already use. Risky actions, like a refund or an email to a customer, wait for a person to approve them, unless one of your rules allows them.
- Works in
- Zendesk, Salesforce, Stripe, GitHub, Slack and Google
- Risky actions
- Wait for a person, unless one of your rules allows them
- Every action
- Written to a log nobody can quietly edit
- Runs
- On your own servers. You can read the code
agentsupport-agent
acts formaya@acme.co
actionstripe.refund
args{ charge: ch_9f2, amount: 250.00 }
- 01Kill switchesOffPASS
- 02ArgumentsValid charge, amount ≤ originalPASS
- 03DelegationMaya may refundPASS
- 04Budgets$250 of $2,000 left todayPASS
- 05RulesRefunds over $100 need an adminHOLD
- 06Risk vs trust
Approved → applied
Approved by dana@acme.co · applied once
evt 4812 #a91c…e07 ← #77b2…1fd
stripe.refund ch_9f2 250.00 · approved · applied
Every action passes the same six checksAllowed · Needs approval · Denied
The shift
Agents are moving from answering to acting
01
Chat
Answers a question. A person copies the answer somewhere.
02
Copilots
Suggests inside one tool. A person still clicks every button.
03
Agents that act
Emails customers, issues refunds, changes records and ships code.
The moment an agent can act, every company needs an answer to two questions: who is it acting for, and who said it could?
The problem
Today's agents work alone, and nobody is watching
01
Keys everywhere
Each agent holds its own API keys to your CRM, billing and email.
02
Nobody approved it
Risky actions happen first and get noticed later.
03
Nothing is shared
Each agent starts from zero. What sales learned never reaches support.
04
Nothing is provable
There is no record anyone can trust of what an agent did, and why.
How it works
One kernel between every agent and everything it touches
People, the agents you build, the agents you already use, your systems and company memory meet in one place, under one set of rules.
People
Approve, delegate and ask
Agents you build
Ready-made by department, or your own
Agents you already use
Claude, Cursor or your own code
OrchKernel
one gate · approvals · budgets · delegation · one memory
Your systems
Salesforce, Zendesk, Stripe, GitHub, Google, Slack, any MCP server or REST API
Company memory
Shared sources and people who opt in, curated
All of them write to one tamper-evident log
Agents propose. Rules decide. The kernel applies.
The rule underneath everything.
See the architectureWhat you can do
Hire agents, build your own, and keep your systems
Set up
Describe your company. It sets itself up.
The Architect turns a description into a blueprint: teams, modules, agents, rules and pages. A person reviews and promotes it. Nothing installs itself.
See how setup worksHire
Agents by department, ready on day one.
Each agent comes with its playbook, tools, rules and a scorecard. Enable a module in one click and map it to your teams.
See the departments- Sales
- Marketing
- Support
- Product
- Engineering
- Founder's office
Build
Build a new agent in an afternoon.
An agent is a record, not a project. It has a plain-language playbook, declared tools, a trigger, a trust tier, a budget and a person it acts for. Changes go through review and a canary that rolls back automatically.
See how agents are builtBring
Bring the agents you already use.
Claude, Cursor or your own code get their own identity and act only through the gate. Each answer is allowed, needs approval or denied, with the reason. Outside agents never hold your vendor keys.
See the actions APIConnect
Keep your systems. Agents work inside them.
Any MCP server, built-in adapters for Zendesk, Salesforce, Stripe, GitHub, Gmail, Google Calendar, Slack and Google Drive, and declarative adapters for any REST API. Records can stay where they are.
See the integrationsPeople in charge
Every action passes one gate
Before anything runs, the gate checks it in order. The answer is allowed, needs approval or denied, with the reason.
- 01
- Kill switches
- 02
- Arguments
- 03
- Delegation
- 04
- Budgets
- 05
- Rules
- 06
- Risk against trust
For example
Support Agent proposes
A refund of $250
Needs approval: a rule says this refund needs an admin
The run waits. The admin sees the exact refund in the inbox.
The admin approves. The run resumes where it waited.
OrchKernel issues the refund exactly once, with its own credentials.
Every step is in the log.
01
Approvals with the exact payload
The approver sees exactly what will happen: the email, the amount, the change. Approve, and the run resumes where it waited.
02
Delegation and trust
An agent never has more than the person it acts for, and revoking it stops everything. Trust is earned: sandboxed, standard, trusted. Budgets cap each agent, team and task.
03
Memory agents can use and can't leak
The Curator files meetings, email, documents and chat from shared sources and people who opt in, and links each item to the customer or deal. New knowledge waits until a person accepts it. Search respects each person's permissions.
04
Proof, and rules as code
Each event is chained to the one before, and
ok audit verify names the first break. Replay any run. Rules, budgets and kill switches live in Git and apply through review.Built for enterprises
Runs where your data lives
- Self-hosted
- One binary and one database file, on your servers or your cloud. Docker and Kubernetes assets are included.
- Models cleared by data class
- Choose your model providers. Each model is cleared for a data class, so restricted data reaches only models cleared for it.
- Sealed secrets, checked egress
- Secrets are sealed at rest. Every outbound request is checked against its allowed hosts.
- Source-available
- Business Source License 1.1. Run it for your own organisation, or one deployment per client. Read the license terms
Put agents on real work. Stay in charge.
Design partners run one real workflow on OrchKernel and get a say in the roadmap.
Source-available · Self-hosted