The questions a risk committee asks. Answered by the system, not by the agent.
Each answer comes from a control Alchemy enforces and a record it writes, not from what a model says about its own work.
| The question | How Alchemy answers it |
|---|---|
| Which agents exist, and what can each one reach? | One manifest is the registry. For every member it states the model, instructions, the tools and actions it may use, the files and memory it may read, and who it may hand work to. Every run is pinned to the exact manifest it started under; changing the rules mid-run requires a recorded amendment. |
| Can an agent exceed its permissions? | Permissions are enforced by the runtime on what an agent actually does, not by instructions in its prompt. A request outside a member's grants is refused and written to the record as a denial. |
| Who approved that action, and what exactly did they approve? | Every decision is recorded with the answer, the account that made it, and a fingerprint of the exact request approved. No agent can approve. |
| What did the agent see before it acted? | Alchemy keeps exactly what each agent was given at every step, so any action can be traced back to the information behind it. |
| Who checked the work before it was called done? | A member that did none of the work opens every deliverable against a written definition of done before the run can finish. On by default. |
| Can an outside action happen twice after a crash? | No. Alchemy logs each action on your systems when it is requested and again when it completes. If something is interrupted in between, Alchemy checks what actually happened instead of trying again, and asks a person when it cannot tell. |
| How much can a run spend? | You set the limits. Per-run ceilings and optional daily and monthly limits are checked before each step, and paid tool calls are reserved before they are sent. Reaching a limit pauses the work for a person. |
| Can we reconstruct a run without the vendor? | Yes. The record lives on your disk, the run's full state can be rebuilt from it, and an export verifies on any machine without Agiliti. |
People decide what matters. Agents do the work in between.
You mark the actions that need a person, such as pushing code, opening a pull request, filing, publishing or a paid call. Alchemy holds each one until an administrator decides, and keeps every role in its lane.
Actions you mark wait for a person
A marked action pauses only the task that asked for it, while the rest of the work carries on. An administrator approves or declines it in the control center or from an alert in Slack, Microsoft Teams or a webhook, and unanswered requests are raised again until someone decides.
The decision names who and what
Each decision records the answer, the account that made it and a fingerprint of the exact request.
No agent can approve
Approval authority belongs to people. Within the team, the manifest names which member may carry out each outside action, so a member that writes a change is not the member that ships it.
An independent check before done
The lead cannot sign off its own work. A member that made none of the deliverables opens every one against the written definition of done, and after repeated failures the run asks a person.
Budgets set before the work starts. Checked before every step.
Finance sets the limits once. Alchemy enforces them across every run, including when many runs start at the same time, and parks the work for a person when a limit is reached.
A dollar ceiling on every run
Each run carries a ceiling on model spend, checked between steps and before each model turn. When the run reaches it, the run pauses and asks a person whether to continue and to what figure.
Paid calls reserved in advance
Declared paid tool calls, such as media generation, are reserved against a limit before they leave your environment, so calls in flight count too.
Daily and monthly limits
Optional daily and monthly limits for each project and for the whole account cover spend across every run, plus what running work may still spend. A new run that would not fit is refused or given a smaller ceiling.
Your key, your bill
Model use is billed by Anthropic to your own API key, at Anthropic's prices. Agiliti meters nothing and marks nothing up.
Your data goes where you send it. None of it goes to Agiliti.
Alchemy runs as one process on a machine you control. The connections that carry your data are the model provider, on your key, and the tools and services you choose to connect.
Claude Code's own telemetry settings are passed through, so you can apply your organization's policy to Anthropic's default reporting on the same machine.
| Destination | What goes | Who controls it |
|---|---|---|
| Model provider (Anthropic) | Prompts, context, tool results and replies for each step | You, under your own agreement with Anthropic and on your own API key |
| Tool servers (MCP) | The arguments of the tools a member is granted, such as a ticket comment | The manifest, which names each server and each tool. A server's tool list is fingerprinted, and any change holds the server until an administrator approves it |
| APIs and integrations | Requests to GitHub, Jira, HTTP APIs, media generation and voice services | The manifest, as declared actions: allowed hosts and request patterns, approvals where you mark them, and a place in the record |
| Alert channels | The run name and the question awaiting a decision | Your alert settings for Slack, Microsoft Teams or webhooks |
| Agiliti | Nothing | Alchemy sends no telemetry, no usage data and no licence checks, and needs no account with us |
Named accounts, clear roles. Secrets kept out of reach.
Administrators
Decide and configureDecide approvals and spend, manage connections and manifests, and delete runs. Every decision in the record names the account that made it.
Users
Run the workStart runs and answer the questions those runs raise. Everyone signs in with a named account.
Agents
Least privilege by defaultStart with no built-in tools. Everything a member can do comes from its grants in the manifest, and it acts only within them.
Network exposure
Local unless you publish itThe control center accepts connections only from its own machine. To give a team access, place it behind your own TLS reverse proxy or tunnel, under your existing network controls.
Sign-in and sessions
Hardened by defaultPasswords hashed with scrypt, strict session cookies with idle and absolute timeouts, sign-in throttling, and default-deny routing on every request.
Secrets
Stored by referenceTokens live in an encrypted credential store, never in the manifest and never shown back. Tool-server sign-in uses OAuth, so Alchemy never sees the password.
Encrypted on your disk. Under a key only you hold.
Encryption at rest
On by defaultAES-256-GCM, record by record, covering run records, each step's context, prompts, transcripts, deliverables, memory, knowledge bases, the credential store and the user store.
Key held outside the data
From your secret managerThe key is supplied from your secret manager or a key file you name, kept apart from the data it protects, and never logged, exported or passed to a process Alchemy starts. Encrypted data on a copied disk, in a backup or in an exported directory is unreadable without it.
Retention you control
Set per environmentRun records are kept until you decide otherwise. Large media files age out on a schedule you set, and a preview shows what a sweep would remove before it runs.
Deletion
Deliberate and accountableOnly an administrator can delete a run, with a stated reason. Alchemy first writes a sealed summary of what was deleted, who deleted it, what it spent and which decisions were made, so the deletion itself stays on the record.
One complete record per run. Verifiable without us.
Alchemy keeps an append-only record of every run: each step's exact context, every declared outside action, every approval, denial and spend decision, and the independent check at the end.
Hash-chained and append-only
Each entry carries the hash of the one before it, so an edited, removed or reordered entry shows up when the chain is checked. The check runs without the encryption key.
Kept by the system, not the agent
Alchemy keeps what each agent is given at every step and logs each action on your systems as it is requested and as it completes. The record reflects what the system did, not the agent's own account of it.
A bundle your auditor can check
Export a run as a self-contained bundle with a single digest over its contents. Keep that digest somewhere independent, and anyone can confirm the bundle is unchanged.
Verifiable without Agiliti
The record is yours, on your disk. Rebuild a run's full state from it, produce a decrypted copy with your key, and verify an export on any machine.
Built to support the frameworks you report against. The EU AI Act first.
Alchemy helps you meet the EU AI Act's requirements for record-keeping and human oversight. The conclusions are yours to draw; the evidence is ready when you need it.
| Requirement | How Alchemy helps |
|---|---|
| Article 12, record-keeping | Automatic, append-only recording of every run: steps and their context, outside actions, approvals and who decided, denials, spend decisions and verification results. |
| Article 14, human oversight | Each agent's reach is written in a manifest a person can read. Marked actions and spend limits pause for a person, who can approve, decline, stop the run or accept the result, and the record names who decided. |
| Article 26, deployer obligations | Logs stay on infrastructure you control, under retention you set, with named administrators responsible for oversight. |
Mapped to the OWASP Top 10 for Agentic Applications.
How Alchemy's controls help you address the risks your security team is already tracking.
| Risk | How Alchemy helps |
|---|---|
| Agent goal hijack | Outside content reaches the model clearly marked as untrusted, and what an agent can do is bounded by its grants whatever it is told. |
| Tool misuse | Per-member grants enforced by the runtime, approvals on the actions you mark, and spend ceilings with paid calls reserved in advance. |
| Identity and privilege abuse | Reach is defined per member, down to the files it may read; credentials are held by reference in an encrypted store. |
| Agentic supply chain | Tool-server tool lists are fingerprinted, and any change is held for an administrator's approval. |
| Unexpected code execution | Agents have no built-in shell. Tests and workspace commands run only as declared actions, confined to the workspace, with a filtered environment. |
| Memory and context poisoning | Memory is off by default. When on, the runtime derives each note's source, and notes with an outside basis are quarantined by default. |
| Insecure inter-agent communication | Members hand off work only along the lines the manifest names, and every hand-off is recorded. |
| Cascading failures | An approval pauses only the task that asked; spend limits pause the run; an independent check stands between the work and done. |
| Human-agent trust exploitation | Approvers see the exact request, and the decision is recorded against it. |
| Rogue agents | Members act only through their grants, cannot approve anything, and cannot run under a changed manifest without a recorded amendment. |
Ready for your security review. And open to researchers.
Your security questionnaire, answered in writing
Send us your questionnaire and we answer it in writing, control by control. A runbook and a threat model come with the software.
Responsible disclosure
Found a vulnerability? Our disclosure policy explains what is in scope and how to reach us.
Questions a security review asks. Answered plainly.
Does Alchemy help with EU AI Act obligations?
Alchemy supports the record-keeping and human-oversight obligations the EU AI Act sets out, with automatic, tamper-evident records of agent activity and human approval of the actions you mark. It supports your compliance work rather than making a system compliant on its own; risk classification and conformity decisions remain with your organization.
How does Alchemy relate to the OWASP Top 10 for Agentic Applications?
Alchemy maps its controls to the OWASP Top 10 for Agentic Applications. Scoped permissions for each member address tool misuse and privilege abuse, human approvals gate high-impact actions, and the tamper-evident record supports investigation and accountability. The mapping shows where Alchemy helps reduce risk; it is not a certification, and your own threat model still applies.
Can an agent approve its own actions or alter the record?
No. Approvals come only from administrators, and no agent, including the lead, can approve a marked action. The record is encrypted and hash-chained, so altering a past entry breaks the chain, and the break shows when the record is verified independently. Oversight stays with people, and the evidence stays intact.
Where is our data stored?
On infrastructure you control. Alchemy is self-hosted and runs as one process on your Windows or Linux machine, so manifests, records and deliverables stay inside your environment. Model requests go to Anthropic on your own API key, under your agreement with them. Nothing is sent to Agiliti, and we need no access to your runs.
How can our security team review Alchemy?
Start with this page, which covers approvals, spend, data flow, identity, encryption, the record and how Alchemy supports the EU AI Act and OWASP guidance for agentic applications. Because Alchemy runs in your environment, your team can inspect it directly, we answer security questionnaires in writing, and the Security page explains how to report a vulnerability.
Bring your risk committee's questions.
Run Alchemy on your own infrastructure in an eight-week design partnership, or start with an evaluation and inspect the controls on your own terms.
