What you get. A working organization, and the proof it is governed.
Six things you keep at the end of week eight, plus the governance pack for your risk committee.
Your organization, in one manifest
Written with you: its members and lead, what each may touch, who hands work to whom, which actions need a person's approval, and what counts as done. Running on your machine, ready to license.
A real piece of your work, handed back
Your goal, delivered as the files your team would open first, checked by an independent member against the definition of done, and approved by your administrator wherever you marked it.
Approvals and ceilings, tuned to your work
Approval rules for every outside action your work touches, a spend ceiling on every task, and daily and monthly limits per project and for the account, set from what your own runs show.
A verifiable record of your first real run
The exported record, with a written account of what each member was given, what it did, what it spent and what waited for a person. Your auditor can verify it independently, without us.
Up to three product changes you ask for
Chosen from what your runs show, and delivered with regression tests in the build you license.
Regression evaluations from your own runs
Your checked runs become test cases, replayed with the same independent check before you change a member, an instruction or a model.
The eight-week plan. Every milestone ends with something you keep.
Week 1: Install and design
We install Alchemy on your machine with you and confirm it with the built-in health check. Then we draft your organization from the role library and the setup wizard, and refine it together: members, hand-offs, approvals and the definition of done. You have: a running installation and your draft manifest.
Week 2: Your first governed run
Ceilings and limits are set before anything runs. Your organization then works end to end against your goal on your own API key, and we review the recorded context of each step together. You have: your first governed run and its record.
Weeks 3 and 4: Connect the real system
We connect the one system your goal must touch, whether your repository, Jira, a tool server or an HTTP API, as declared actions that wait for a person wherever you want one. We tune ceilings and limits from what the record shows. You have: the live connection, tuned ceilings and an inventory of every outside action.
Weeks 5 and 6: The changes you asked for
Up to three product changes, chosen from what your runs exposed, delivered with regression tests. We review every tool your organization can call and agree with you, in writing, which ones become approval-gated actions. You have: the changes, their tests and a signed-off action inventory.
Week 7: The account of your first real run
We write up your first real run from its exported record: what each member was given, what it did, what it spent and what waited for a person. You have: a written account your auditor can check against the record.
Week 8: Handover
We walk your CISO and risk committee through the governance pack, hand over the export and your regression cases, and agree the licence. You have: the finished work, the governance pack, the export and your regression suite.
The governance pack. Built for your risk committee.
Everything your CISO needs to answer two questions by handing over a file: who approved that action, and what did the agent see before it acted.
Threat model
Fitted to your manifest: what each member can reach, what it cannot, and the controls that enforce it.
Data-flow diagram
What leaves the machine, and only as you configure it: model calls to Anthropic on your key, the systems you connect and the alert channels you switch on. Nothing goes to Agiliti. Records stay encrypted on your disk.
Retention and deletion
Retention for generated media set to your policy. Records are kept; deleting a run leaves a sealed tombstone of what it did.
EU AI Act mapping
Articles 12, 14 and 26, on record-keeping, human oversight and deployer obligations, mapped to what a run records, the approval flow and the manifest.
Action inventory
Every outside action your organization can take: which member may request it, whether it needs approval, whether it can be undone and what it costs.
Sample record and reading guide
One real run from your partnership, exported, with a short guide to reading it, recomputing spend per member and verifying the chain.
What we need from you. Four things.
An Anthropic API key
Billed to your own keyModel use bills your key directly. We never see it, proxy it or mark it up.
One machine
Windows or LinuxAlchemy runs as one self-hosted process on Windows or Linux with Node.js 22.
One goal
Real and boundedA change your team would review as a pull request, a file your reviewers would sign, or a cut your post team would deliver.
An administrator
Approves what you markThe person who decides the actions you mark for approval, from the control center or from an alert. Often your CISO or the accountable lead for the work.
Commercial terms. In plain words.
What you pay us
$5,000, fixed, for eight weeks. Nothing is metered by us.
What you pay Anthropic
Your own model use, on your own API key, never marked up.
What you sign
A short design-partnership agreement to run Alchemy on your infrastructure for eight weeks.
What you keep
Your manifest, every record, the written account of your first real run, the governance pack and your regression cases.
What happens after
The $5,000 is credited in full to a first-year licence, quoted per organization.
Questions before you apply
Can we build on the agents we already have?
Yes, as an Alchemy organization. The roles, instructions and repositories your agents use carry over as members, so approvals, ceilings and the independent check apply to everything they do. We set this up with you in the first weeks.
Does any of our data reach Agiliti?
No. Alchemy runs on your infrastructure and sends nothing to Agiliti. Model calls go to Anthropic on your own key.
Who can approve an action?
Only the administrators you name. No agent can approve, and every decision is recorded with who made it and exactly what it allowed.
What makes a good first goal?
Real work with a clear owner and a clear finish: something a named person would otherwise review and sign.
Apply for a design partnership
We read every application and reply from contact@agiliti.ai. We use what you write here only to answer you.