Skip to content
github-actions[bot] edited this page Jul 24, 2026 · 3 revisions

agentgear wiki

A Rust library and derive macro for shipping a coding-agent plugin from a binary. It gives the binary a canonical install/uninstall/update/self-heal lifecycle for 25 agents. Claude Code and copilot-cli get the full plugin flow: a setup subcommand replaces the user typing /plugin marketplace add + /plugin install (or copilot plugin marketplace add + plugin install on copilot's own CLI). The other 23 config-merge backends merge into their own config files instead.

The README is the overview. These pages hold the reference.

Mental model

The crate lives in your binary, not in the plugin tree. The binary embeds the plugin at compile time as a compressed blob and hands it to one or more agent backends at runtime. 25 backends ship today: Claude Code, copilot-cli, and 23 config-merge agents; the full list is on the Agent backends page.

your binary  ──derive──▶  PluginHost  ──reconcile──▶  AgentBackend
                                                        │
                     claude / copilot-cli ──▶ the tool's own plugin CLI (marketplace + materialize)
                     23 config-merge agents (codex, cursor, gemini, …) ──▶ each tool's own config file

embeds plugin/ (.tar.br blob) ── materialize ──▶ ~/.local/share/<name>/current@<client>   (claude + copilot-cli, each its own <client>)

Pages

page topic
Getting started add the crate, the derive, the build guard, hook wiring
Plugin tree what plugin/ holds, what each backend can translate out of it
How it works lifecycle to CLI mapping, materialize, the self-heal state table
Agent backends the unsealed AgentBackend trait, install models, adding your own
Harness comparison harness-first at-a-glance: support grid, config paths, scopes
Capabilities capability-first depth: a page each for MCP, hooks, commands, agents, skills, instructions + detection, scopes, native ingestion
Testing your host hermetic runs against a temp HOME, reading back what each backend wrote
Doctor the health report: shared checks, per-agent checks, fix hints
Types and errors the programmatic API: Outcome, AgentReport, Capabilities, the components IR, Error

Status

v1 ships 25 agent backends: Claude Code + copilot-cli (full plugin lifecycle) plus 23 config-merge backends. Every non-CC backend was verified against its real shipping binary, with positive and negative controls before any parse was trusted. 19 per-tool docker legs (the GUI/IDE and no-surface backends have none) run in CI as a matrix, one job per tool. Hermetic and unit tests are green in plain cargo test. The AgentBackend trait is unsealed.

Linux is CI-gated. Windows was probed by hand against Claude Code 2.1.201: a junction created without elevation, with validate, marketplace add, install, and list --json all resolving through current; the no-BOM and no-trailing-whitespace path rules in materialize come from that same run. macOS is a first-class target in the design but is not exercised on a real host in CI. Every CI job runs on ubuntu-latest.

Clone this wiki locally