Terminology
wraptool’s documentation uses one name per concept. This page is that list.
Read it as a glossary if you are new. If you are writing documentation, treat the Preferred term column as binding and see the style rules in Contributing.
The actors
The three most-confused words on this site. They are not synonyms and never substitute for each other.
| Term | Means | Not |
|---|---|---|
| agent | The AI acting through a harness: it decides what to do and calls MCP tools | The program it runs in; the model weights |
| harness | The CLI program or IDE integration the agent runs in — claude, opencode, gemini, cursor, roo, pi, agy. Version-locked in the harness pool |
The AI itself; a container; a wrapper script |
| model | The language model behind the agent. Use only when the point is genuinely about the model | The agent; the harness |
Write agent, or coding agent on first use in a page. Do not write “the AI”, “the assistant”, “the coding assistant”, or “the AI harness” — the last conflates two different things.
The boundary
| Term | Means |
|---|---|
| policy | The rules in the configuration’s tools: block that decide which operations are allowed. The user-facing name for what wraptool enforces |
| privileged operation | An operation needing credentials, network access, or host authority the agent never holds. git pull is the canonical example |
| mediation | wraptool’s model: the agent requests an operation, wraptool executes it. Contrast with a credential proxy, where the agent holds and uses a credential |
| capability | One allowed operation as seen by the agent, exposed as an MCP tool |
| approval | An operator’s decision on a capability request, made in the review UI |
Write policy, not “whitelist”, “ruleset”, or “access rules” in prose. whitelist survives as a Go package name and in identifiers; it is not the word for the concept.
Write privileged operation. Do not alternate with “protected operation” or “sensitive operation”. Use credentialed only where the point is specifically about credentials rather than about privilege in general.
The processes
| Term | Means |
|---|---|
| server | The wraptool MCP server process. Agent-facing. Started by wraptool serve, or by wraptool up on your behalf |
| review UI | The operator-facing web interface for approving and denying capability requests. Host-facing, loopback only, never agent-facing |
| runtime | The container the harness runs in: the Guix runtime or the Dev Container runtime |
Write review UI. Do not write “Web UI”, “admin UI”, or “admin listener” — the first is vague about who it is for, and the other two suggest a general administration console rather than a request-review surface. Where the point is the network listener rather than the interface, write the review UI’s listener.
The harness pool
| Term | Means |
|---|---|
| harness pool | The version-locked store of harness installations shared across projects. pool alone is fine after first use on a page |
| lock | The recorded resolution of a harness to exact versions and integrity hashes |
| resolve | Turn a harness request into a lock. Needs the network |
| realize | Build a generation from an existing lock. Offline |
| generation | One immutable realized harness installation. Retained so harness rollback can return to it |
| receipt | The record proving a generation was realized from its lock |
| quarantine | The waiting period before a newly published version becomes eligible |
| provision | Run resolve or realize through a backend — Guix, or Docker/Podman via the OCI provisioner |
Conventions
- wraptool is lowercase, including at the start of a sentence. It names both the project and the executable.
- Command names are code-formatted and written as invoked:
wraptool up,wraptool harness install. - Harness names are code-formatted and lowercase:
claude,agy. - Say configuration for the concept and
config.yamlfor the file. Avoid “config” as a noun in prose; keep it in commands and paths where it is literal.
For writers: preferred and avoided
| Preferred | Avoid |
|---|---|
| agent, coding agent | the AI, the assistant, coding assistant, AI harness |
| agent-facing | AI-facing |
| harness | wrapper, agent binary, tool |
| policy | whitelist, ruleset, access rules |
| privileged operation | protected operation, sensitive operation |
| review UI | Web UI, admin UI, admin listener |
| harness pool, pool | the harness store, the cache |
| configuration | config (as a prose noun) |
Adding a term is a documentation change like any other: put it in this table in the same commit that introduces it.