AI tool guides

Choose the page named after the AI tool you use. Each guide explains which onemind client supports it, how to connect it, where its instruction file lives, and how to verify the result.

For the CLI, a detected repository is not active until you explicitly bind its exact target to an ordinary Project in the setup flow. Complete that activation before onemind sync; detection alone grants no write or upload permission. Native BETA users instead follow Projects and folders, which prepares compatible connected targets and leaves uploads disabled.

New instruction-sync targets are Project-scoped. Codex, Antigravity, Claude Code, OpenClaw, OpenCode, Hermes Agent, Pi, Grok Build, DeepSeek Harness, Cursor, Qoder, TRAE, and CodeBuddy connect to an explicit local Project directory. The current server composes the Team-role Project Mind followed by the selected ordinary Project's Project-role Mind once, and the CLI writes that atomic package to the Project's tool-native instruction file. Both are the same project Mind kind; stored AI-target Mind resolution is not yet part of this runtime. Existing global targets remain visible only for explicit migration and are not offered for new setup. Cursor supports a CLI-selected Team Project target and an independently activated ordinary Project repository target. Qoder and TRAE are additional CLI repository targets. CodeBuddy uses the compatibility target key workbuddy: discovery can suggest candidate paths, but only an exact Project binding activated by the user can sync.

Native BETA Project sync also maintains the tool-specific stable user-wide bootstrap described in Mind Sync. This adapter configuration is separate from legacy global Team-content bindings and never contains Project content. CLI target registration and native folder binding have separate setup paths.

When a Codex, Claude Code, or Grok Build task requests a tool permission while running, the device owner can answer in Matrix or the macOS app without pausing the task. The Matrix composer also accepts / slash commands for active Codex, Claude Code, and Grok Build sessions. In a Claude Code session, @ lists Claude Code subagents; Codex and Grok Build have none.

Capability integrations

Instruction sync and the AI tool integration are separate. Sync writes the reviewed Mind into the target's instruction file. The integration bundle adds the onemind skill, authenticated MCP connection, and native host tools where the host supports them.

Use the explicit onemind integrations commands (install, update, repair, uninstall) to change each integration. The official installer only stores the signed bundle; only these explicit commands change a host integration. Integration actions never install, update, or remove Project skills. The Codex integration action changes only the bundled Codex marketplace/plugin; it never runs the full onemind installer, replaces the CLI or app, restarts the service, or opens an app.

For another operating system, download https://onemind.team/downloads/integrations/vVERSION/OneMindIntegrations-VERSION.zip together with its .sha256 and .sigstore.json files. onemind verify-release-integration VERSION OneMindIntegrations-VERSION.zip OneMindIntegrations-VERSION.zip.sigstore.json performs read-only publisher verification. onemind import-signed-integration-release VERSION OneMindIntegrations-VERSION.zip OneMindIntegrations-VERSION.zip.sigstore.json ABSOLUTE_STORE authenticates the exact bytes and publishes an immutable generation with retained witnesses. Host installation is then performed only by onemind integrations install TARGET; installers and shell scripts never mutate host plugins directly. The .sha256 file detects corruption but does not authenticate the publisher.

The official installer stores the exact verified bundle before any host lifecycle action. Inspect it and explicitly choose one target at a time:

onemind integrations list
onemind integrations status codex
onemind integrations install codex
onemind integrations update codex
onemind integrations uninstall codex

list --json and status TARGET --json separate package state from host state. onemind never searches the current Project PATH; use --host-executable with one trusted absolute path for a nonstandard host location. Official CodeBuddy 2.131.0 fresh-profile plugin list, plugin validate, and model execution produced no usable output within 60 seconds; its logs repeated missing user-settings errors, exhausted file watchers, and disabled hot reload. Passive status never launches an unowned profile. When the fixed ~/.codebuddy profile is absent, explicit install returns requires_host_initialization without launching CodeBuddy; an existing profile retains the bounded 15-second inventory preflight. Cursor uses a native receipt-bound local package and reports connected only after a fresh cursor-agent JSON marker probe; missing or unauthenticated hosts remain offline-only/unverified. TRAE remains package-prepared/manual, Mind-only targets reject local mutation, and every mutation requires one explicit target without changing Project skills. Experimental Windows wrapper code remains restricted to the exact allowlisted native executables and npm cmd-shim templates documented below.

Evidence scope: Codex 0.146.0-alpha.3.1, Claude Code 2.1.220, OpenClaw 2026.7.1-2, Pi 0.82.0, and AGY 1.1.9 completed their exact install/list/update-or-reinstall/uninstall lifecycles against this bundle in isolated homes. The shared OpenClaw/Pi package identity is @onemind/onemind; OpenClaw loads the exact manifest ID onemind without a package-name mismatch warning. Qoder 1.1.5 completed the exact native lifecycle, but authenticated vendor-artifact CI is not claimed. With an audit-only PYTHONPATH, the retained Hermes 0.19 runtime completed discovery, enable, disable, remove, and clean reinstall; this is still not evidence of a clean installed host. Vendor remove retained the disabled/override configuration, and reinstall plus enable cleared the disabled set. Fresh Claude inventory writes ~/.claude.json and a backup; absent OpenClaw inspection writes SQLite state; passive status therefore never launches an unowned host. Claude's wrapper and native bytes matched the official npm integrity and its isolated lifecycle passed. The native binary reported Team ID Q6L2SF6YDW, but strict codesign/spctl validation failed; this remains a vendor-signature investigation rather than a onemind lifecycle failure, and onemind makes no Apple-signature trust claim. The official AGY 1.1.9 installer completed against its published manifest, but this run did not independently re-hash the compressed archive; the extracted binary reported Team ID EQHXZ8M8AV, while strict Apple authority/notarization verification was unavailable, so onemind makes no Apple trust claim. CodeBuddy commands remain source-confirmed only because fresh inventory exceeded the onboarding budget. Do not generalize these exact evidence categories into broader version support.

After installing or updating an integration, start a new tool session. Authentication remains a separate on-use step: approve the onemind OAuth connection in hosts that expose MCP authentication.

When an AI tool connects to the onemind MCP server, you choose which permissions to grant. Each permission covers a specific set of operations; grant only what the tool needs for your workflow. For authorization steps, see Sign-in and accounts > Authorize the CLI or an AI tool.

None of these operations write local files directly; file writes require a separate approved workflow on the destination device.

For direct Codex host-manager installation from the extracted marketplace root, the validated syntax is:

codex plugin marketplace add "$ROOT" --json
codex plugin add onemind --marketplace onemind-app --json

The plugin specifier is onemind with a separate --marketplace onemind-app option. Do not use the unvalidated onemind@onemind-app form.