Solutions · IT
You run it. You approve it.Straight answers for both.
Access reviews, joiner-mover-leaver, the service desk queue, the runbook only one person can run. IT gets more out of an agent platform than most teams, and it is also the team that has to say yes first. This page covers both.
Runbooks that run · deny-by-default reach · SAML 2.0 single sign-on
- Runbooks
- Markdown in your repo, not in a head
- Reach
- Granted per agent, never inherited
- Every action
- Allowed, gated, or blocked
- Deployment
- Cloud, your VPC, or on-prem
01The handoff
The queue, and the runbook only one person can run.
IT work fits this unusually well. It is procedural, it is already written down somewhere, and the cost of skipping it stays invisible until an audit or an incident.
Job 01
Joiner, mover, leaver
The mover is where it pays off. A joiner gets a checklist and a leaver gets urgency, but someone who changes team quietly keeps both sets of access forever. A scheduled session finds those people and drafts the revocations.
Job 02
The access review that actually gets done
Who has what, for how long, and when they last used it. It produces the review with a proposed action on every row, not a spreadsheet that gets forwarded twice and approved unread.
Job 03
Service desk triage and a first draft
It reads the ticket, classifies it, checks whether your own docs already answer it, and drafts the reply with the article linked. The ones it cannot answer get routed with the diagnosis already attached.
Job 04
The runbook that lives in one person’s head
Write it down as a skill file and a session can run it. That is the lasting version of "ask Marta": it works while Marta is on holiday, and improving it is a change request, not a corrected memory.
Job 05
Licence and seat reconciliation
What you pay for, against what is assigned, against what was opened this quarter. It writes the reclaim list per tool, with the last-used date on every row.
Job 06
Configuration drift, reported, not fixed
It compares what your docs say is configured with what actually is, and reports the difference. Reporting drift is safe and useful. Quietly correcting it is neither, so it does not.
02What comes back
A review with a recommendation on every row.
What IT needs is not a dashboard. It is a list with a defensible reason on each line, ready so that approving it actually does something.
| Principal | Grant | Granted | Last used | Recommendation |
|---|---|---|---|---|
| j.okafor | db · read-write | 14 mo | 9 mo | Revoke: moved to Support in Q1 |
| deploy-bot | registry · push | 8 mo | 2 h | Keep: in the release path |
| a.lindqvist | admin console | 22 mo | never | Revoke: granted for a migration |
| support-agent | helpdesk · read | 3 mo | 11 m | Keep: scoped, read only |
Agents are principals too
Machine has a real account, member, group and role model, with per-resource permissions for people and for agents. An access review that covers your people but not your automation is half a review.
A recommendation, with the reason on the row
Revoke because they changed team. Keep because it is in the release path. You can approve a review in ten minutes when the reasoning sits on the line, not in a separate document.
It reports. It does not quietly fix
Revoking is a write, and writes are where you set a gate. A review session should be set up to read everything, change nothing, and propose.
03Where it reaches
Your whole estate, reached the way it really is.
IT tooling is the least uniform stack in the company. Connect what is in the catalogue and define the rest yourself. Either way the credential is decrypted on our side and never enters the machine.
Google Workspace and Microsoft 365
Accounts, groups, mailboxes and calendars: what most joiner-mover-leaver work runs against. Read and write are separate actions, each with its own setting.
GitHub
Repo and org membership is access too, and it is the access most likely to outlive its reason. It reads the org state and reports what does not match your documented roles.
Your service desk
The mainstream helpdesk products are in the Easy connect catalogue behind their own OAuth screens. Read the queue, draft the reply, and update the ticket where you allow it.
Everything with an API and no catalogue entry
That is most of an IT estate. Point Machine at an OpenAPI or Postman description, a GraphQL endpoint, a plain HTTP base URL, or a remote MCP server. It reads the source, works out the auth, and turns each operation into a tool with its own Allow, Ask or Block.
A machine you already run
A computer connector opens a tunnel to a host you control, so a session can reach something that was never going to face the internet, without putting that host on the public network.
Easy connect covers 3,000+ apps through their own OAuth screens. We will not claim your identity provider, your MDM and your network vendor are all in it. For an IT estate the direct connector types are usually the real path, and they take about the same three minutes.
04How it runs
Answer the ticket now. Run the review every quarter.
The same session, started three ways. IT gains the most from the scheduled mode, because the work that gets skipped is the work with no deadline.
01 · On demand
The question in the help channel
Someone mentions the bot in the IT channel. The thread becomes a session, the session reads your docs, and the answer lands back in the same thread with the source quoted.
02 · Human-assisted
It stops before it changes access
Set every write against an identity or an entitlement to Ask. The run holds at that call with the exact change in front of you, and picks up from there when you approve. Set the irreversible ones to Block instead.
03 · Automated
The review nobody schedules, scheduled
A cron trigger opens the quarterly access review and the monthly licence reconciliation. Each trigger names the agent it runs as, so the unattended session has exactly the reach the attended one has. No more.
05Control
Now you review us.
You are also the team that has to approve this. Here are the answers, including the ones that are less flattering than what you usually read on a page like this.
01
One isolated machine per session
Every session runs on its own disposable Linux machine, on its own branch. Sessions never share a filesystem or a working tree, and the machine is thrown away afterwards. What we will not tell you is that egress is controlled at the network. Nothing does that, so assume a sandbox has outbound access.
02
Approval gates are off by default
The shipped default is permissive: an action runs unless you said otherwise. For you, this is the most important sentence on the page. Set Ask and Block explicitly, per action or with a pattern rule, and note that a rule can match on the arguments, not just the tool name.
03
Merge is default-deny
Work reaches main through a change request. An agent cannot merge unless an admin grants project.cr.merge in your project config. That grant is text in a versioned file, so widening it is itself a reviewed change, not a setting that quietly moved.
04
Two kinds of credential, and they behave differently
A connector credential is brokered server-side and never enters the sandbox. A runtime secret you grant a session IS a real environment value inside it, readable by any command the agent runs. Both are the right design for their job. Do not let anyone tell you the second one is hidden from the model.
05
Identity, audit, and where it runs
Single sign-on is SAML 2.0. Every tool call is written to an audit record with the agent, the person or trigger, the outcome and the approver. Versioned configuration makes policy changes visible and reversible. For an isolated setup, talk to us.
06Other teams
Same machine. Every other team.
One project, one set of connectors, one memory that keeps growing. Each team writes the skills for its own work. Nobody builds a second system.
- SalesResearch, drafts and CRM cleanup, held for your yesOpen
- MarketingProduction work in your voice, because the voice is a fileOpen
- ProductFeedback turned into specs, with the evidence attachedOpen
- EngineeringReproduce it, patch it, open the change requestOpen
- FinanceThe close, the reconciliation and the variance note, with the working shownOpen
- PeopleScheduling, kits and onboarding, never the hiring decisionOpen
- Data ScienceA real machine, a real query, an analysis you can re-runOpen
- All solutionsEvery team on one page, plus the rules that never change between them.Open
IT
Read the security pagebefore the pitch.
One governed platform for IT agents. Any model, your keys. Every action recorded.
- One machine per session
- Lands as a change request
- Every action recorded