Plugin types
Skills and packages guide the runtime; they are not executable tools. A component or MCP-backed tool performs the actual operation.
Lifecycle
1
Create the definition
Provide a clear name, description, capability summary, type-specific configuration, and availability settings.
2
Define safety metadata
Classify mutation intent, target scope, reversibility, side effects, approval mode, timeout, and output limits. These declarations must match real behavior.
3
Test connectivity
Test provider configuration without enabling the plugin for normal runtime use. For MCP, testing authenticates and initializes the server but does not replace the synchronized tool catalog.
4
Synchronize capabilities
Sync component or MCP metadata. RevoEngine stores the discovered capability state so operators can inspect drift and keep the last known good MCP manifest when a refresh fails.
5
Attach and validate
Attach the plugin to an Assistant thread, Agent configuration, Agent policy, or individual run. Start with a read-only validation before enabling writes.
6
Operate the lifecycle
Update, re-sync, disable, restore, or delete the plugin as its provider changes. Soft-deleted definitions preserve historical metadata.
Availability and attachment
Plugins can be attached to instance defaults, an Assistant thread, an Agent definition or policy, or an individual run. Discovery settings can make an active plugin available for appropriate work, but discovery is not an authorization grant. Disabled, deleted, or policy-blocked plugins cannot execute. Keep plugin availability narrow. Attach only the capabilities needed for the responsibility and validate read and write behavior separately.Component tools
A component tool maps a stable plugin operation to a versioned RevoEngine component. Define a concise input schema and output schema, keep execution bounded, and expose only the component operations the AI should call. Use component tools when business behavior already lives in RevoEngine or should share the platform’s normal versioning, execution, and observability.MCP servers
MCP plugins keep authentication in RevoEngine Secrets and synchronize the remote capability contract for operator review. Each admitted operation retains its own safety and approval policy, so read and write capabilities from the same server can be governed differently. When an MCP refresh fails, inspect the plugin status and last synchronization error. Re-test authentication and review the remote contract before enabling or trusting changed operations.Skills and packages
A plugin skill should describe one focused workflow, its success conditions, preferred tools, and safety boundaries. It can depend on exact plugin capabilities without duplicating their credentials or schemas. A package is the top-level aggregation boundary for a coherent offering. Keep dependencies acyclic and shallow, and avoid nesting packages inside packages. This makes review and lifecycle management predictable.Safety model
Each executable plugin operation can declare:- read, write, destructive, or external mutation intent;
- reversible, operator-reversible, irreversible, or unknown reversibility;
- Assistant-thread, Agent, tenant, or external target scope;
- no, conditional, or always-required approval;
- a safety tier and human-readable side-effect summary;
- timeout, result size, and enablement limits.
Operational checklist
- Use the narrowest input and output schemas.
- Test and sync before attachment.
- Verify that mutation intent and reversibility match the provider.
- Require approval for external writes and destructive actions.
- Set bounded timeouts and output sizes.
- Treat a changed MCP manifest as a contract change.
- Disable a plugin that no longer has a trustworthy current capability state.
- Re-run a read-only Assistant or Agent validation after each material update.
Tools and skills
See how plugin capabilities join the normal runtime.
Security and governance
Review identity, roles, secrets, and audit boundaries.