DEPLOYMENT PLANNER
Plan your deployment. Know the cost.
Choose the services your team needs. Understand infrastructure, responsibilities and costs before you deploy.
You pay AI WorkDeck for licensing and deployment services. You purchase and pay for servers, GPUs, storage, model APIs and third-party services yourself. They appear here only to estimate your overall investment and are excluded from our service fees.
Team & workload
User count determines licensing; concurrency, documents and models determine infrastructure. Presets also fill an editable monthly server allowance.
Where models run
Storing data privately and running models privately are separate decisions.
Cloud model usage assumptions
All rates are CNY. Defaults are budgeting placeholders, not prices for a particular model. Replace them with current supplier rates.
Monthly cost = users × (input × input rate + output × output rate) × cloud share
Choose supporting services
Basic file storage, database and built-in Git are in deployment scope. Choose optional services below; you supply external accounts and quotas.
Default OCR/transcription rates are editable assumptions too. Blank means unquoted; 0 means existing quota or no incremental spend. Unknown costs are never presented as free.
Infrastructure & operations
Planning guidance only, not a load-tested capacity promise. Disk starts at 100 GB + 5 GB/user, rounded up; backups at 2×. Adjust for documents, Git history and retention. Model GPUs are separate.
Base environment: Linux, JRE 21, PostgreSQL, HTTPS / internal DNS, persistent disks and backup destination. Validate GPU memory and throughput for model weights, quantization, context and concurrency before purchasing.
What each service must support
Checked against application source and deployment files. An implementation is not acceptance in your environment. This checklist separates core features, API procurement and adaptation.
Application & databaseRequired
Linux, JRE 21, Nginx / HTTPS and PostgreSQL. Vector search also needs pgvector and an embedding model. Desktop workbench or Office/WPS add-in connects to private services.
Acceptance: Verify account isolation, client connections and restore. A full browser-only workbench is outside standard scope.
Models & per-user meteringRequired for AI
Existing OpenRouter / platform pathways record user, project, session, model and input/output tokens. Private model gateways need integration.
Acceptance: Test streaming, multi-turn tools, structured arguments, usage receipts, long context and optional vision; verify user attribution, retry deduplication, quotas and reconciliation.
Files & backupsRequired
Project files use a persistent filesystem; Git history adds storage. Back up the database, project/bare repositories, service state and required encryption keys.
Acceptance: Test permissions, disk alerts and restore. Validate NAS mounts; generic S3 / MinIO is not a drop-in replacement. Keep a backup in a separate failure domain.
Git & team file versionsBuilt in
JGit smart-HTTP, project repositories, device tokens and membership are built in. Basic use needs no separate GitLab / Gitea purchase.
Acceptance: Test multi-user pull/push, access control, conflicts and repository restore. Source hosting is separate from business-file collaboration.
Identity, licensing & networkConfirm before delivery
Official desktop builds still depend on account and online license checks. Confirm private identity, client endpoints, update sources and offline licensing. SMS/email depends on the login design.
Acceptance: Agree outbound allowlists, HTTPS/internal DNS and license lifecycle. Full isolation requires replacing account, model, search and update pathways.
Four model-service acceptance checks
- Attribute input/output tokens to a user, model, project and request; missing usage is not zero usage.
- Test streaming, cancellation/timeouts, tool calls and multi-turn results; validate long documents and vision on real acceptance examples.
- Verify quotas, concurrency, retries and duplicate-charge protection; reconcile supplier billing separately from local statistics.
- Vector retrieval also needs embeddings; include primary, auxiliary, embedding and image models in cost and outbound inventories.
Clarify responsibilities as well as hardware
| Role | Customer responsibility | Deployment deliverable |
|---|---|---|
| Business owner | Name one owner for user counts, scope and acceptance examples. | Requirements, scope and acceptance record. |
| IT / operations owner | Name one owner, shared or contracted, for networks, backups and monitoring. | Single-environment installation, permissions, restore demonstration and handover. |
| Model / API owner | Provide service accounts, budgets and model/data entitlements. | Integration and usage checks; separate assessment for private models. |
| Security / acceptance owner | Approve outbound data, retention, access roles and recovery targets. | Agreed-scenario acceptance; HA, certification and specialist audits need separate scope. |
Private deployment offer
One organization, one environment, internal use. Priced by total users with access; over 200 users requires a quote.
CNY reference prices exclude tax. Renewal estimates use the current same-scope annual offer. Final quotes confirm tax, versions, scope and acceptance.
| Users | License / year | Setup / once |
|---|---|---|
| ≤ 20 | ¥9,800 | ¥6,000 |
| ≤ 50 | ¥19,800 | ¥10,000 |
| ≤ 100 | ¥39,800 | ¥16,000 |
| ≤ 200 | ¥49,800 | ¥24,000 |
Standard deployment includes: Existing components in one environment, basic API integration, one restore demonstration, one admin training and 30 days of delivery-issue follow-up.
On-site work, ongoing operations, migration, custom development, HA and model adaptation are scoped separately; specialist services may be estimated at ¥2,000 / person-day . Infrastructure, equipment and external services remain the customer’s responsibility.
AGPLv3 use and commercial licensing are separate paths. This page does not require every self-hosted user to buy a commercial license. See the commercial license terms。
Reviewed 2026-10-13 against the application repository. Usage rates are editable assumptions. Generation models, SMS/email, data licenses and unlisted integrations must be scoped separately; unquoted does not mean free.