Glossary
A short definition for each term, with a link to the page that covers it in depth.
Core terms
Section titled “Core terms”Assistant A named persona with a model, a set of skills, and an avatar. The unit you talk to. See Assistants.
Team Several assistants working together on one job under a leader. See Teams and Standing Companies.
Standing Company A team you keep between sessions, rather than spinning up for one task. See Teams and Standing Companies.
Agent A reusable definition (a system prompt plus limits like turn count) that an assistant or the engine can adopt. See Agents.
Skill A named prompt snippet with variables, optional shell expansion, and per-skill model or permission overrides. See Skills.
Workflow (ForgeFlow) A declarative, multi-stage agent pipeline written in RON. See Workflows and the RON reference.
Channel An inbound and outbound messaging surface that connects an assistant to a chat network such as Telegram or Slack. See Channels.
Constitution The always-loaded rules document, reloaded fresh every turn, with optional per-specialist overlays. See The Constitution.
Engine terms
Section titled “Engine terms”Tool A capability the agent can call. Eight are built in (Read, Write, Edit, Bash, Grep, Glob, Delegate, ToolSearch); MCP servers add more. See Tools and the Tool Registry.
MCP Model Context Protocol. The standard the engine uses, as a client, to connect to external tool servers over stdio, SSE, or streamable-http. See MCP.
Memory Per-project storage that the engine indexes across sessions so the agent can recall earlier context. The desktop surfaces it as the Memory page and Wiki. See Memory and the Wiki.
Wiki The desktop knowledge view built on top of memory: concepts, topics, and a knowledge graph. See Memory and the Wiki.
Project A scoped working context that groups conversations and files. See Projects and Workspaces.
Profile
A named bundle of provider and model settings you switch between with --profile, with extends for inheritance. See Profiles.
Compaction How the engine keeps long sessions within token budget: three automatic tiers (microcompact, autocompact, emergency) plus output compaction. See Context Compression.
Plan Mode A read-only exploration mode for designing a plan before the agent acts. See Plan Mode.
Swarm A worktree-isolated set of workers the engine can dispatch across a repository. See Swarm.
ACP Agent Client Protocol. The transport the desktop app uses to drive the embedded engine. See ACP Server and Client.
JSON stream
The protocol a host program uses to drive the engine over --json-stream, exchanging command, event, and approval messages. See JSON Stream Protocol.
wcore-evolve The self-improvement system: a GEPA-style loop that evolves skills, with a deterministic scorer (no model in the scoring step) as its trust boundary. See Self-Improvement (wcore-evolve).