Supported harnesses

HarnessTap registers 44 agent harnesses today. Each harness declares which resource types it can scan, compose in plugins, and materialize on disk, plus default project and global paths. The registry in src/platforms/registry.ts is the source of truth; ht harness list prints the same set at runtime.

For portability caveats (hooks with ${*_PLUGIN_ROOT}, Claude Code mods, OpenCode server plugins, instruction-only skill emission, and mirror warnings), see Portability limits.

Resource types

HarnessTap separates context-side resources (what the model sees) from environment-side resources (how it runs).

Context-side material resources

These are the atomic resources you scan, curate, attach to plugins, and re-emit during apply:

Plugin composition adds two attachment kinds on top of material resources:

Environment resources

Environments carry how values that override matching plugin resources during apply (home → plugin default → deck active):

Only harnesses whose registry entry includes env_vars, model_config, and/or permissions and whose native serializer implements those surfaces receive environment values on disk. Today that means Claude Code and Codex for the full environment trio.

All harnesses whose serializer emits MCP config still receive MCP ${VAR} substitution from the environment cascade at apply time (tokens in MCP env / args). OAuth sessions managed by the host IDE are outside this path: see Environments: MCP authentication limitations.

Plugin manifest layouts

scan and plugin-source import recognize these on-disk plugin trees (in addition to per-harness project files):

Official Claude plugin cache trees at ~/.claude/plugins/cache/<marketplace>/<plugin>/<version> are also valid scan/sync roots when they have no plugin.json (LSP-only plugins such as gopls-lsp, whose lspServers live in the marketplace catalog). Name and version are taken from the cache path. If that directory is a marketplace checkout, import uses the nested plugins/<name> tree. Marketplace source / path values like ./plugins/gopls-lsp resolve from the marketplace repo root, not from .claude-plugin/.

Scanning a repo with both harness files and a plugin manifest merges both sources automatically. See Portability limits: dual-mode scan.

Plugin install and sync providers

During apply, HarnessTap can install and sync plugins from host install trees for:

Claude Code and Copilot CLI home scan also import installed marketplace plugins as plugin_pin library resources (pins only; use resource sync to materialize children). Claude reads ~/.claude/plugins/installed_plugins.json. Cursor home scan inventories ~/.cursor/plugins/local/ (including HarnessTap copies, which keep name@marketplace in .harnesstap-plugin.json) plus cache/marketplace trees referenced by Cursor install records: not every cache directory, not HarnessTap host-plugin cache, not ~/.claude/plugins/cache. Copilot CLI reads ~/.copilot/installed-plugins/<marketplace>/<plugin>/ when ~/.copilot/settings.json or that install tree is present.

ht harness sync (global) copies those install trees onto peer harnesses that have a native plugin root: Claude ←→ Cursor. Claude additionally records installed_plugins.json and enabledPlugins in ~/.claude/settings.json. Cursor copies land in ~/.cursor/plugins/local/ as real directories (not symlinks) and are replaced on the next sync. Cursor does not load ~/.claude/plugins/. Project apply does not write that user-wide folder. Claude mods (in-process hooks.json + JS/TS) copy only onto platforms with hostPluginRuntimeModules (Claude Code); Cursor still receives the portable rest of the pin. Desktop lists Claude plugin/skill trees on the Cursor harness only while Cursor still loads third-party Plugins, Skills, and other configs; when that setting is off, Harnesses does not attribute ~/.claude/ or ~/.agents/ to Cursor. When OpenCode (or another harness that does not load those plugin roots) is in the active set, the same sync also materializes plugin skills into ~/.agents/skills/ and agents/commands into that harness’s native folders. harnessSync.pluginResources in config.jsonc selects symlink (default), copy, or clone (copy-on-write, falls back to copy).

Other harnesses still benefit from plugin-source scan and resource sync when you point at an install tree or plugin repo, but do not have an automated install provider yet.

Claude plugin pins and marketplace metadata (plugin show → claude block) apply only when materializing to claude-code.

Cursor home scan inventories every newest cache version directory under ~/.cursor/plugins/cache/, plus marketplaces/ and local/, including when Cursor has that plugin turned off. It does not read state.vscdb. HarnessTap host-plugin cache and ~/.claude/plugins/cache are still not Cursor plugins.

Serializers

Filter native harnesses at the CLI:

ht harness list --supported
ht harness list --format json

Skill emission modes

Most harnesses emit skills to native skill directories. These harnesses declare skillEmission: instruction-only: skills merge into instructions or rules instead of a separate skill tree:

Cursor additionally honors project cursor_skill_mode (agents-skills default, agent-requested, always-on). Default apply writes skills under .agents/skills/ only, not also as .cursor/rules/*.mdc. Inspect with ht harness project status --project . --format json.

Agent / subagent bridging

Harnesses with a delegated subagent model normalize into a shared canonical shape (description, instruction body, model, reasoning_effort, sandbox/readonly flags) via src/services/agent-bridge.ts:

Plugin import scans agents/*.md and agents/*.toml (Codex packs). Cross-harness apply re-emits valid native files per target harness.

Harness reference

Legend for the Resources column: instr instructions · skill skills · rule rules · mcp MCP · perm permissions · hook hooks · agent agents · cmd commands · env env vars · model model config.

On-disk paths (selected harnesses)

These are the primary project paths HarnessTap scans and writes. Global paths follow the same keys under each harness home directory (see ht harness list --format json).

Claude Code user MCP lives in ~/.claude.json (mcpServers at the top level). Global apply merge-overlays that key and leaves OAuth session and projects (local-scope MCP, trust) untouched. Project MCP remains .mcp.json. Local MCP (projects[<absPath>].mcpServers) is scanned for inventory with source ~/.claude.json#local:<absPath> and is never applied. See Portability limits: Claude Code MCP scopes.

Cursor global user skills live under ~/.cursor/skills/. Cursor also maintains app-managed built-ins under ~/.cursor/skills-cursor/: HarnessTap inventories those on profile status / apply-preview (host_managed.cursor) and on Desktop Harnesses (app-managed panel) but never imports, applies, discards, or removes them from disk. Cursor plugin installs live under ~/.cursor/plugins/ and are listed only when Cursor's own install records say they are installed. When third-party Plugins/Skills/configs are enabled, Desktop Harnesses also lists related trees Cursor can read: Claude ~/.claude/plugins/ and ~/.claude/skills/, plus the shared ~/.agents/skills/ hub. When that setting is off, those foreign trees are omitted from the Cursor harness.

Copilot CLI home state lives under ~/.copilot/. Detection uses ~/.copilot/settings.json and ~/.copilot/installed-plugins/ in addition to ~/.copilot/skills/ and ~/.copilot/mcp-config.json. Installed plugins are imported as pins from ~/.copilot/installed-plugins/<marketplace>/<plugin>/.

Generic harnesses in the skills-only tier use harness-specific skill roots such as .kilocode/skills/, .crush/skills/, or .factory/skills/ with AGENTS.md instructions: see the registry for the full list.

Cursor plugin directories under ~/.cursor/plugins/cache, marketplaces/, and local/ count as installed, including when Cursor has the plugin turned off. plugin-* MCP folders count as installed too. A folder is matched to the plugin named in its manifest or under ~/.cursor/plugins/cache/<marketplace>/<plugin>/<version>, and that plugin is not listed a second time. Status does not read state.vscdb or the Cursor logs folder. A HarnessTap database write that changes a profile is included in the next status. Adding a library resource that no profile uses does not refresh status. Whether Cursor loads other harness trees is read from settings.json only, and defaults to on.

Goose context engineering

HarnessTap maps Goose context engineering surfaces as follows:

Antigravity notes

Antigravity IDE, Antigravity CLI (agy), and AGY share the same workspace layout under .agents/. Workflows (.agents/workflows/*.md) map to HarnessTap commands. Global config lives under ~/.gemini/ (GEMINI.md, config/mcp_config.json, skills/, config/global_workflows/). This is distinct from the thinner gemini-cli entry.

OpenCode notes

OpenCode discovers skills from its native trees plus Claude- and Agents-compatible directories. HarnessTap inventories those compatibility paths and scans them, but apply still writes only the native OpenCode locations.

Grok Build notes

Grok Build’s native layout under .grok/ maps as follows:

DeepSeek Harness notes

DeepSeek Harness is a developer preview; the Cordis patch schema can change. Home files live under $DSH_HOME if set, otherwise ~/.dsh. Detection uses distinctive .dsh/skills/ or .dsh/hooks/: not AGENTS.md alone.

Muse Code notes

Muse Code (Meta) user settings live at ~/.config/muse/settings.json. The file must include "schema_version": 1 whenever it exists. HarnessTap merges mcp_servers and user hooks into that file and leaves model defaults, TUI, runtime_capabilities, telemetry, and managed_hooks_path alone. MCP is user-settings only: there is no documented project MCP path.

Detection uses distinctive .muse/hooks.json and/or ~/.config/muse/ (settings.json or the config directory). Shared AGENTS.md / .agents/skills/ alone do not count as Muse. muse init only writes AGENTS.md; a Muse-only machine is still detected from the home config directory.

MiniMax Code notes

MiniMax Code (mcode, npm @minimax-ai/code) stores user data under ~/.minimax (profile: ~/.minimax-<name>; override MINIMAX_DATA_DIR / MAVIS_DATA_DIR). The installer directory ~/.minimax-code is not user data and is not used for detection. Runtime settings live in config.yaml; HarnessTap does not rewrite that file. MCP lives in mcp.json (mcpServers).

Detection uses distinctive .minimax/skills/, .minimax-plugin/, and/or ~/.minimax/ (or legacy ~/.mavis/). Shared AGENTS.md / .agents/skills/ alone do not count as MiniMax. mcode init . only writes AGENTS.md; a MiniMax-only machine is still detected from the home data directory.

Related commands

ht harness list                          # all registered harnesses
ht harness list --supported              # native serializers only
ht harness status                        # global main + alias selection
ht harness project status --project .    # per-project harness prefs + cursor_skill_mode
ht scan . --dry-run              # detect harnesses and resources in a repo
ht environment list                      # named environments (env vars, models, permissions)

See also

• Portability limits: fidelity matrix and workarounds
• CLI command reference: full command surface