Glossary
This page gathers the terms used across the docs, each with a short plain-language definition and a link to the page that explains it in depth. Terms are grouped by related topic.
How it works
| Term | Meaning | Learn more |
|---|---|---|
| Agent | The "does the work" assistant Silicon Worker acts as when running a task: instead of just replying, it calls tools to get the job done step by step. | How the agent works |
| ReAct / agent loop | The repeating "plan → act → check" work loop: decide the next step, call tools to do it, check the result, then start another round. | How the agent works |
| Turn | One pass through the loop: a single model call plus the tool calls it triggers in that step. Complex tasks usually take several turns. | How the agent works |
| Compaction | When the conversation grows past what the model can handle at once, earlier work is automatically summarized into a more compact form to free up room to continue — usually with no action from you. | How the agent works |
| Subtask / dispatch | The lead agent breaks a complex task into smaller subtasks, dispatches them to different experts or child assistants in parallel, then gathers the results. | Experts & teams |
Capabilities & extensions
| Term | Meaning | Learn more |
|---|---|---|
| Tool | A capability the agent uses to act — reading and writing files, running commands, operating the browser and apps, and so on; every tool carries a risk level. | Built-in Tools · Tools & risk |
| Skill (SKILL.md) | A reusable instruction document (SKILL.md) capturing the standard procedure for a type of task, so Silicon Worker handles it to the same standard each time. | Skills · Using skills |
| Plugin | An extension unit that packages skills, experts, hooks, and MCP servers together; installing it brings in those capabilities in one go. | Using plugins |
| Expert | A saved, reusable role defined for a type of work, typically holding role instructions, bound skills, and available tools; it can have its own private memory. | Experts & teams |
| Team | A combination of several experts plus their shared skills, letting multiple roles collaborate on one task; a team definition can itself be packaged and shared. | Experts & teams |
| MCP | An open protocol for bringing external tools and services in as tools the agent can call; Silicon Worker acts as an MCP client connecting to external services. | Connect MCP |
Memory & knowledge
| Term | Meaning | Learn more |
|---|---|---|
| Memory (Hermes) | Silicon Worker's long-term memory (internally named Hermes), which remembers your preferences, procedures, and project context across sessions and gets to know you over time. | Memory · Manage memory |
| Private memory | Memory owned by a specific expert, taking effect only in that role's related work. | Memory |
| Global memory | Memory shared across roles, suited to things you want remembered in any situation, such as how to address you, your usual language, and general habits. | Memory |
| Knowledge base | A collection of material you import and can search, bindable to a session, expert, or team for on-demand lookup during tasks — complementary to memory. | Knowledge base |
Safety & permissions
| Term | Meaning | Learn more |
|---|---|---|
| Risk level (Safe / Low / High) | The label each tool (even different uses of the same tool) carries: Safe is read-only with no impact, Low has limited impact, High may cause significant or irreversible changes. | Tools & risk · Built-in Tools |
| Permission mode (Free / Controlled / Restricted) | A single setting for how much Silicon Worker asks first: Free asks only for high-risk actions, Controlled asks for any risky action, Restricted denies by default and allows only what you explicitly permit. | Permission modes · Tools & risk |
| System permissions | The OS-level grants needed to reach system capabilities (accessibility, automation, full disk access, calendars, reminders, notifications, and so on), granted on demand. | System Permissions |
| Workspace / sandbox | A per-session working directory that file operations and command execution are confined to, with no access beyond it. | Tools & risk |
Models & running
| Term | Meaning | Learn more |
|---|---|---|
| Provider | A source of large-model service, connected through a unified gateway; most use the OpenAI-compatible protocol, while the Claude series uses the Anthropic native protocol. | Model Providers · Configure providers |
| Primary/backup model & failover (fallback) | A session can set a primary model plus one or more backups; when the primary hits a transient problem and retries still fail, it automatically switches to a backup to continue. | Model Providers |
| Scheduled task | A task that fires automatically on a schedule, one-off or recurring; it runs on your machine, so the app must stay running and the device powered on and online. | Scheduled tasks & remote |
| Remote IM connection | Connecting a chat tool so you can send tasks and receive results remotely from your phone or another device; available connectors depend on your version. | Scheduled tasks & remote |