-
Notifications
You must be signed in to change notification settings - Fork 0
Harness Comparison
Side-by-side reference for the 25 backends agentgear ships: Claude Code, copilot-cli, and 23 config-merge agents. Every cell was checked against the real shipping binary (or the tool's shipped extension source, for the IDE-bound ones) on 2026-07-16 (copilot-cli's native rewrite 2026-07-18), with positive and negative controls before any parse was trusted.
"Translated" means agentgear read-modify-writes the tool's own config file so the plugin's
components land in that tool's native shape. Claude Code and copilot-cli are the exception: both
get the full plugin lifecycle through their own CLI (claude plugin / copilot plugin), no
config merge. A backend writes only when it detects the tool installed and the tool has a surface
at the target scope.
What each backend translates. — means the surface is skipped, with the reason. claude and
copilot-cli are plugin-native (the whole tree ships through the tool's own plugin CLI, not a
per-surface translation); every column reads native for both.
| harness | mcp | hooks | commands | agents | skills |
|---|---|---|---|---|---|
| claude | native | native | native | native | native |
| copilot-cli | native | native | native | native | native |
| cursor | ✓ | ✓ | ✓ | ✓ | ✓ |
| devin | ✓ | ✓ | ✓ | ✓ | ✓ |
| qwen-code | ✓ | ✓ | ✓ | ✓ | ✓ |
| droid | ✓ | ✓ | ✓ | ✓ | ✓ |
| codex | ✓ | ✓ (inert until trusted in /hooks) |
✓ | ✓ | — (not translated) |
| gemini | ✓ | ✓ (a CC tool-name matcher never matches gemini's own) | ✓ | ✓ | — (not translated) |
| augment | ✓ | ✓ | ✓ | ✓ | — (not translated) |
| crush | ✓ | ✓ | ✓ (TUI-only, not loaded by crush run) |
— (no file-writable surface, issue #1807) | ✓ |
| kilo | ✓ | — (no hooks key in the schema) | ✓ | ✓ | ✓ |
| vscode-copilot | ✓ | ✓ (CC matcher dropped) | — (out of scope) | ✓ | — (out of scope) |
| cline | ✓ | ✓ (CC matcher dropped) | ✓ (workflows) | — (surface unimplemented) | — (surface unimplemented) |
| opencode | ✓ | — (JS/TS plugin API only) | ✓ | ✓ | — (not translated) |
| omp | ✓ | — (TS extension API only) | ✓ | ✓ | — (not translated) |
| kimi | ✓ | ✓ | — (slash-command analog is a skill) | — (built-in sub-agents only) | ✓ |
| goose | ✓ | ✓ | — | — | ✓ |
| kiro | ✓ | — (kiro hosts hooks only inside user-owned agent configs; no target agentgear can own) | — | — | ✓ |
| antigravity-cli | ✓ | ✓ (a CC tool-name matcher never matches agy's own) |
— | — | — |
| zed | ✓ | — | — | — | ✓ |
| openclaw | ✓ | — (JS/TS module hooks only) | — | — | ✓ |
| jetbrains-copilot | ✓ | — | — | — | — |
| antigravity | ✓ | — | — | — | — |
| amp | ✓ | — | — | — | — |
| pi | — (no native mcp surface) | — | — | — | — |
Every ✓ above carries one filter: ${CLAUDE_PLUGIN_ROOT} expands only inside Claude Code's own
runtime, so a config-merge backend skips any mcp server whose command or args carry the token,
and any hook whose command does, rather than write an entry that would launch the literal path. The
canonical plugin-shipped server puts the token in args
({"command": "node", "args": ["${CLAUDE_PLUGIN_ROOT}/server.js"]}), so that shape is the one that
gets dropped everywhere except Claude Code. A ✓ says the surface translates; it does not promise
every entry in it survives. claude and copilot-cli never run this filter, since both copy the
tree verbatim: Claude Code expands the token in its own runtime, and whether copilot's raw-shape
copy of hooks/hooks.json fires at all is unconfirmed.
Two more surfaces sit outside these five, both declared by the host rather than shipped in the tree, so neither rides the components IR:
-
instructions (
PluginHost::instructions, always-loaded guidance text). Onlyopencodewrites it today, a dedicated file registered in itsinstructions[]array;claudereceives the same text through the MCPinitialize.instructionsfield instead. -
status line (
PluginHost::statusline). Five backends write it into their tool's own single slot:claude,copilot-cli(user scope only),qwen-code,antigravity-cli(user scope only),droid. That slot holds one value, so agentgear stashes whatever it displaces and restores it on uninstall. Full rule: Status line.
The user-scope config each backend writes. A row names the file (or dir) carrying mcp, and a second path whenever that backend's hooks, commands, agents, or skills land under a different root. Most backends honor a home/config-dir env override; the notable ones are in Agent backends. Project scope, where a backend serves it, sits under the working tree.
| harness | user config file |
|---|---|
| claude |
<config>/plugins/ (via claude plugin; CLAUDE_CONFIG_DIR); <config>/settings.json for a declared status line |
| copilot-cli |
~/.copilot/installed-plugins/<mkt>/<plugin>/ (via copilot plugin; native, whole tree copied; COPILOT_HOME); ~/.copilot/settings.json for a declared status line |
| amp | ~/.config/amp/settings.json |
| antigravity | ~/.gemini/config/mcp_config.json |
| antigravity-cli |
~/.gemini/config/ (mcp_config.json, hooks.json); ~/.gemini/antigravity-cli/settings.json for a declared status line |
| augment | ~/.augment/settings.json |
| cline |
~/.cline/data/settings/cline_mcp_settings.json (mcp); ~/Documents/Cline/ (hooks, workflows) |
| codex | ~/.codex/config.toml |
| crush | ~/.config/crush/crush.json |
| cursor | ~/.cursor/mcp.json |
| devin | ~/.config/devin/config.json |
| droid |
~/.factory/ (mcp.json, hooks.json, commands/, droids/, settings.json for a declared status line) |
| gemini | ~/.gemini/settings.json |
| goose |
~/.config/goose/config.yaml (mcp); ~/.agents/plugins/<plugin>/ (hooks, skills) |
| jetbrains-copilot | <config>/github-copilot/intellij/mcp.json |
| kilo | ~/.config/kilo/kilo.json |
| kimi |
~/.kimi-code/mcp.json + config.toml
|
| kiro | ~/.kiro/settings/mcp.json |
| omp | ~/.omp/agent/mcp.json |
| openclaw | ~/.openclaw/openclaw.json |
| opencode | ~/.config/opencode/opencode.json |
| pi | none (detect-only) |
| qwen-code | ~/.qwen/settings.json |
| vscode-copilot | project only: <repo>/.vscode/mcp.json + .github/
|
| zed |
~/.config/zed/settings.json (mcp); ~/.agents/skills/ (skills) |
Which scopes each backend serves. Scope::User targets the home-level config,
Scope::Project { path } that directory's. A backend with no surface at the requested scope is
skipped: no write, no marker, no error, so a project-scope install lands only on the backends
marked ✓ under project. Self-heal runs at user scope only.
vscode-copilot is the one project-only backend. Its user config is rejected on purpose: a
non-default VS Code profile relocates User/mcp.json; the active profile is VS Code internal state
that cannot be read off disk.
| harness | user | project | notes |
|---|---|---|---|
| claude | ✓ | ✓ | the project path reaches the CLI as its working directory |
| copilot-cli | ✓ | — | the copilot CLI takes no --scope; a native install is user-level. project .mcp.json is real, unwritten |
| amp | ✓ | — |
.amp/ needs amp mcp approve <name> and is scanned at the exact cwd, no walk-up |
| antigravity | ✓ | — | |
| antigravity-cli | ✓ | ✓ | project scope loads only after the folder is trusted |
| augment | ✓ | — | the tool serves .augment/ too; the backend stays user-only |
| cline | ✓ | ✓ | mcp is user-global whichever scope you pass; hooks and workflows honor both |
| codex | ✓ | ✓ | project needs trust_level="trusted" in the user config.toml. codex mcp list never reads a project config |
| crush | ✓ | ✓ | a same-key entry in $XDG_DATA_HOME/crush/crush.json silently outranks ours |
| cursor | ✓ | ✓ | project wins a same-key mcp collision |
| devin | ✓ | ✓ | project overrides user on a same name |
| droid | ✓ | ✓ | droid reads a third tier between them, any ancestor's .factory/mcp.json, outside our write surface |
| gemini | ✓ | ✓ | an untrusted folder suppresses mcp and commands at both scopes together; project hooks are blocked, user hooks are not |
| goose | ✓ | — | no project extensions file |
| jetbrains-copilot | ✓ | — |
<project>/.github/mcp.json is real, unwritten |
| kilo | ✓ | ✓ | |
| kimi | ✓ | — |
<cwd>/.kimi-code is a real native mcp scope, undeclared because hooks have no project surface |
| kiro | ✓ | ✓ | |
| omp | ✓ | ✓ |
OMP_PROFILE/PI_PROFILE relocate user scope to ~/.omp/profiles/<name>/agent, which the backend does not track |
| openclaw | ✓ | — | one user-global file regardless of cwd |
| opencode | ✓ | ✓ | |
| pi | ✓ | — | detect-only, nothing written at either scope |
| qwen-code | ✓ | ✓ | project hooks need folder trust; project mcp, commands and agents apply unconditionally |
| vscode-copilot | — | ✓ | |
| zed | ✓ | ✓ | the project write is unconditional; trust_all_worktrees gates only whether zed auto-starts project mcp servers |
stdio is the verified path on every backend; remote (http/sse) fidelity varies per tool. The full per-harness verdict table now lives on the capability page: MCP § remote fidelity.
Several tools read Claude Code's own config or plugin trees directly — the exact tree agentgear materializes — so a translation is redundant and agentgear retires it. The full who-reads-what table and the retirement rules moved to Native ingestion.
- Capabilities: the capability-first view — one page per surface, with the deep per-harness tables (MCP shapes + fidelity, hook events, native ingestion).
- Agent backends: the trait mechanics and adding your own backend.
- How it works: the Claude Code lifecycle, materialize, self-heal.