Your own organization. Running on your work, on your infrastructure.
Alchemy runs the organization you describe. Start from one of the example setups below, or from a blank manifest, and shape it around your own work.
Start from an example setup, or from the setup wizard
Begin with the example below that is closest to your job and point it at your own repository, tickets or documents. Or describe what the work should produce, and the setup wizard drafts a validated organization from the role library, matched to what your machine can reach.
Mark what needs a person, and set the ceilings
Declare each outside action and mark the ones that wait for approval. Set a ceiling per task, and daily or monthly limits where you want them. It all lives in one manifest, with no code to write.
Evaluate it, then put it to work
Evaluation access comes under a short written agreement. Your organization runs on your infrastructure and your own Anthropic API key, and a design partnership turns it into the one you licence.
Ship changes with their evidence. Every outside step approved by a person.
Two software jobs that end where your release process already expects them: a reviewed pull request, backed by a record your approver can read in minutes.
A code change, delivered as a change record plus a pull request
You hand it: an issue in your repository and the rules for who may touch what. What comes back: the change record, which you open first, with the review verdict, the separation-of-duties audit and the test results beside it, plus a pull request on GitHub. The member that writes the code, the member that writes the tests and the reviewer each work only within the folders and tools their role allows, and the auditors that certify the evidence cannot read the repository at all. Pushing the branch and opening the pull request are separate actions, and each waits for an administrator's approval.
A Jira ticket, taken to a signed-off release
You hand it: a Jira ticket key. What comes back: the UAT pack, which you open first, backed by vetted requirements, test cases, a delivery plan, the build and a clean run of the full test suite. When a ticket is too vague to build and test as written, the team drafts a short list of focused questions for the person who raised it, and that comment waits for approval before it posts. The developer and the test writer work in separate spaces. Approving the push is the UAT sign-off; the pull request and the move to Done follow from it.
Files prepared for the person who signs. Your reviewers decide; the team does the preparation.
Three jobs where the deliverable carries a person's name. Alchemy drafts, cites and assembles the file. The determination stays with your accountable official.
Export control: a release memo prepared for your Empowered Official
You hand it: the technical-data documents proposed for release and your reviewer roles. What comes back: the release memo, which you open first, with a drafted jurisdiction and classification for each document, quoting the public control text it relies on, and an access matrix for every reviewer role. Access to controlled documents is enforced at the tool, so a role that stands in for a non-U.S. person cannot open them, and every refusal is on the record. The release itself waits for approval.
AI-use disclosure: an audit response built on the client's consent
You hand it: an advertiser's audit request covering AI use in a campaign, and the client's executed AI rider. What comes back: the audit response, which you open first, with a consent record, an asset register, a per-asset EU AI Act Article 50 determination and the disclosure copy, drafted for a person to assert. Consent comes first: without the executed rider, the team issues a refusal that names the missing document, and stops.
Pharma promotional review: a reference pack tied to the label
You hand it: a campaign brief and the approved product label. What comes back: the HCP email and its annotated reference pack, which you open first, with every promotional claim tied to a numbered section of the label and quoted exactly. The reviewer’s verdict comes with it. Filing the pack waits for an administrator's approval.
Generated media you can account for. Every shot traced to the generation that made it.
When a client, a platform or a regulator asks how a piece was made, the answer is already assembled, shot by shot.
Post-production with shot-level provenance
You hand it: a creative brief and an allowance for paid generations. What comes back: the finished cut and its disclosure pack, which you open first. For every shot, the pack records the model, the prompt as sent, the provider's job reference and a fingerprint of the file that came back. Each paid generation is reserved against the allowance before it is sent. A provenance auditor with no access to any generation provider measures the cut itself, and a reconciler that made none of it matches both accounts shot by shot. Delivery waits for approval, then copies the film into a folder you own.
Questions about fit. Answered plainly.
What can an Alchemy agent organization deliver?
Any job you can describe as an organization with a definition of done. The example setups include a Jira ticket taken to a pull request and a signed-off release, a code change delivered as a change record and a pull request, an export-control release memo, an AI-use disclosure pack, a pharma promotional-review reference pack and media with shot-level provenance.
Can Alchemy take a Jira ticket to a pull request?
Yes. The team reads the ticket, sends focused questions back to its creator when it is unclear, vets the requirements, writes test cases, plans and builds the change, and runs the full test suite into a UAT pack. The delivery manager’s approval of the push is the UAT sign-off; then the pull request opens and the ticket moves to Done.
Is Alchemy suited to regulated or compliance-heavy work?
Yes. It is built for work where evidence matters. The example setups include an export-control release memo, an AI-use disclosure pack and a pharma promotional-review reference pack. Each deliverable is checked against a written definition of done, with human approvals and a verifiable record behind it, while the final judgment stays with your experts.
How does Alchemy handle AI-generated media?
Image, video and voice generation run as governed actions, with each generation counted against its paid-call ceiling and traced to the shot it belongs to. The disclosure that comes with the cut accounts for every generated asset, and delivery waits for a person’s approval, so producers and legal reviewers can see how each shot was made.
Can we build our own setup?
Yes. The example setups are starting points, not limits. Any work you can describe as an organization, with members, a lead, permitted actions, approvals and a definition of done, can be written as a manifest. A design partnership is the fastest way to shape your first organization together with our team, on your own infrastructure.
Bring the job you need signed off.
Eight weeks on your infrastructure, with your administrators approving what matters and a verifiable record of every step.