Repository navigation
deployment product capabilities
Prepared for desktop release v1.25.135. Deployment mode is explicit; a loopback address, local login or SSO configuration does not determine installed products.
| Installation | Default product set |
|---|---|
| Source local / DMG | Goals/workflows, Crew |
Source local with --with-server-products
|
Local products plus Code, Relays, Brain, Vault |
| Server | Configured server products and account permissions |
Local users use their own CLI for standalone coding. Workflow/Crew builders retain their native CLI providers and workflow script authoring.
Normal local installs retain project MCPs, project secrets, local knowledge and workflow learnings. The packaged desktop does not include Code or the four shared server products or support the source-development opt-in.
agent_go/pkg/productpolicy intersects:
-
AGENTWORKS_DEPLOYMENT_MODE=local|server. -
AGENTWORKS_LOCAL_SERVER_PRODUCTS=1, set by the source opt-in. -
AGENTWORKS_ENABLED_PRODUCT_SURFACES, when explicitly configured. -
AGENT_PRODUCTS, retaining the existing shared-server Brain/Vault baseline unless an explicit surface selection excludes them. Dedicated SparkQuill deployments retain their own allowlist.
An old local surface list containing only excluded server products falls back to the normal local product set. An account can narrow this policy and cannot enable an uninstalled product. Product availability remains separate from project grants, token scopes, Vault group/tool/regex permissions and Brain folder roles. Existing external Brain tokens retain their folder-authorized server access even without a Brain app-home entitlement.
Disabling Vault also removes shared secret names/values from runtime selection; it does not fall back to unrestricted legacy global-secret access. A project copy with the same name still follows its existing project permissions and precedence.
- Resolved product profiles and client-visible feature metadata omit disabled tools and skills without modifying the canonical manifest registry.
- Agent wrappers and workflow base agents project system instructions, trusted builtin skills and supporting references, tool descriptions and schemas. Workflow steps, schedules and delegated agents use these shared constructors.
- Native bridges advertise admitted tools. Bound executors recheck current availability, so retaining a tool definition does not grant access.
- Attached and fallback skill reads filter product-owned skill names. Known
disabled skill projections carrying
.agentworks-managedare removed before a CLI starts. User-authored/unmarked folders and symlinks remain intact. - External platform MCP catalogs and guidance use the installation policy; handler-specific authorization continues to apply.
Trusted optional text uses <!-- product:<id> -->…<!-- /product --> blocks.
Only explicit blocks are projected; arbitrary user wording is not removed by
keyword. Imported project skills and learnings are not rewritten. Keep new
platform tool/skill ownership declarations in productpolicy current when
adding product capabilities.
Previously migrated shared knowledge remains marked as shared when Brain is disabled. The agent receives an unavailable-product explanation, and the retired local archive stays blocked. No secret/knowledge migration is performed here.
Regression tests cover ordinary local, source opt-in, shared server, explicit server allowlists, canonical Workflow/Crew/Code guidance, resumed definition assembly, cached execution, step skill resolution, external MCP, safe skill cleanup and existing reader/bot/project grants. Product navigation/Ctrl+K and desktop packaging tests cover the local product set. These checks use no paid model requests.
Restart the backend to load the implementation. For a changed installation profile, restart its native CLI conversations as well. Saved chat history remains historical context; it is not rewritten. No deployment to hosted servers or model requests are required by this change.
Auto-synced from docs/ on main. Edit there, not here.