uygnoey Planning a government deployment
Claude Code and Claude Cowork are available in public beta through Claude for Government Desktop, delivered through a FedRAMP High authorized environment. This skill turns that announcement into the decisions and evidence a rollout actually requires.
Two properties drive most of the planning and are easy to miss: conversation history is stored locally on the agency-managed device, and usage exports contain metering data only. The first moves the endpoint into scope. The second is what lets you answer an ATO or IG question without moving sensitive material.
Instructions
1. Fix the authorization boundary first
Record where processing happens and where data rests, because every later decision depends on it.
- Inference runs inside a FedRAMP High authorized environment.
- Conversation history is retained locally, on the agency-managed device. In the system security plan this is agency-held data under agency endpoint policy — do not describe it as vendor-retained.
- Agencies receive capabilities on the same release schedule as commercial customers. Treat this as both a benefit and a change-management obligation.
Anything your program needs that the announcement does not state is an open question to confirm, not a gap to fill by assumption.
2. Collect the evidence
Three documents exist, at two levels of availability:
| Document | How to get it |
|---|---|
| FedRAMP Secure Configuration Guide | Public |
| Formal change notification | Public |
| Penetration-test summary | Under NDA, through Anthropic's trust center |
Map each to the reviewer question it answers before the assessment meeting rather than during it. Use templates/evidence-map.md.
3. Design delegated administration before provisioning anyone
A department-level administrator allocates seats to sub-agencies, and the limits ride on identity rather than on per-component configuration.
- Connect the identity provider and define groups so one group maps to one sub-agency.
- Set per-group rate limits, dollar caps, and allowed models through SCIM group mappings.
- Use layered configuration to set sub-agency defaults for what Claude can connect to and which features are available.
- Test deprovisioning explicitly: removing a user from the group must remove access.
Getting the group structure right first is what lets a department authorize once and delegate operation, instead of repeating procurement and configuration per component.
4. Bound the spend
Consumption-based pricing and appropriated funds reconcile through one mechanism: usage is purchased in fixed increments with a hard not-to-exceed cap. Say that plainly when finance asks what prevents an open-ended bill.
- Choose standard seats or a customized tier with spend and model limits.
- Size the purchase increment against expected usage and set the cap.
- Route automatic burndown alerts to someone with authority to replenish. An alert that reaches no one who can act is not a control.
- Verify that per-user and per-model tracking in the admin console produces what your financial reporting needs.
5. Set up oversight you will actually run
- Assign named reviewers and a cadence for the hash-chained audit log, reviewable directly in the product by organization administrators.
- Record that sensitive operations on Anthropic's side require two-person approval — this is the answer to "what prevents unilateral vendor-side action."
- Document the usage-export procedure and note that exports are metering data only.
6. Deploy through MDM and treat the endpoint as in scope
The desktop application deploys through standard agency MDM platforms. Because conversation history is local, endpoint policy has to cover disk encryption, backup handling, retention, and device loss and wipe. Pilot with one ring before broad distribution.
7. Contract and record maturity
New customers request access at claude.com/solutions/government. Anthropic remains the
contracted billing party; no separate cloud provider relationship is required. The offering
is announced as public beta — put that in the risk register rather than discovering it during
review.
Working checklist across all seven steps: templates/rollout-checklist.md. Full control surface as announced: references/controls-inventory.md.
Examples
A security reviewer asks where the data goes
"If someone pastes internal code into this, where does it end up?"
Two answers, and both matter. Inference runs inside a FedRAMP High authorized environment. The conversation history is stored locally on the agency-managed device — which means the retention question is answered by your own endpoint policy, not by a vendor retention schedule. Reviewers who expect a vendor-side answer to the second half are usually satisfied faster once the scope shift is made explicit.
Finance objects to consumption pricing
"We can't sign something with an unbounded bill against appropriated funds."
Usage is purchased in fixed increments with a hard not-to-exceed cap, with per-user and per-model tracking in the admin console and automatic burndown alerts before the balance runs low. The cap is the control; the alerts are what keep the cap from becoming an outage. Name the alert recipient in the rollout record.
A department needs different limits per component
"Three sub-agencies, three different risk postures, one authorization."
Map each sub-agency to an identity group, then use SCIM group mappings to set that group's rate limits, dollar caps, and allowed models, and layered configuration to set its defaults for connections and available features. Seats are allocated by the department-level administrator. The department authorizes once; the components differ by configuration rather than by separate agreements.
An IG asks who used what
"We need usage figures for the quarter, and we can't disclose case material."
Usage exports are metering data only, which is precisely why they can answer the question. Pair the export with the hash-chained audit log for administrative actions. Both are available to organization administrators without moving sensitive material out of the boundary.