Tedix plugin

Use it

• Claude chat and Cowork: add a custom connector named Tedix with the URL https://connect.mcp.tedix.dev/mcp, then sign in.
• Claude Code: claude plugin marketplace add tedix-hq/tedix-plugins, then claude plugin install tedix@tedix-plugins --scope user, then sign in via /mcp.
• Codex and the ChatGPT desktop app: codex features enable mcp_2026_07_28, codex plugin marketplace add tedix-hq/tedix-plugins, codex plugin add tedix@tedix-plugins, then codex mcp login tedix; in the ChatGPT desktop app, enable tedix in the Plugins Directory.

The human walkthrough is at https://docs.tedix.dev/learning-paths/first-plugin; the published plugin and marketplace is https://github.com/tedix-hq/tedix-plugins.

Cloud package and first use

For Claude chat, Cowork and Claude Code, see the Claude setup guide. --host claude builds the native Claude format from the same skills; --local adds opt-in context hooks. An explicit --mcp-url selects a self-hosted or local gateway without changing CLI bindings.

Build a reproducible review ZIP from this source with Bun:

bun packages/cli/scripts/package-plugin.ts /tmp/tedix-public.zip
bun packages/cli/scripts/package-plugin.ts --local /tmp/tedix-local.zip

For the default OpenAI artifact, root plugin.json owns the canonical tedix identity. The builder imports review/cases.json and review/release-notes.md as review expectations, not recorded passes. The current plugin version is the version field in plugin.json; the host reports the installed one. The default cloud artifact contains the manifest, portable MCP connection, skills and assets. --local adds the opt-in hooks/hooks.json and host compatibility files to the same plugin identity. The hooks contain no scripts: each runs tedix hooks <name> in the installed Tedix CLI, which is their only requirement, and exits silently when that CLI is not installed. Neither artifact contains app registrations or credentials. The local hooks need Tedix CLI 0.2.0 or later. Building a ZIP does not install, submit or publish it.

OpenAI supports MCP, skills and trusted local hooks in one plugin. Omitting hooks from the cloud artifact is a Tedix execution-environment choice, not an OpenAI prohibition: a web installation cannot supply the separately installed and authenticated Tedix CLI that runs the hooks. Review changed hook definitions in the local host before enabling them. No second plugin is needed; the only recorder is the opt-in decision-capture hook described in the Codex guide.

Portable mcp.json uses the Agent Plugins streamable-http schema. The host-native .mcp.json declares the same remote server with type http and the URL https://connect.mcp.tedix.dev/mcp, and nothing else. Neither file lists scopes or holds credentials: the host's OAuth flow, browser consent and server enforcement determine access.

tedix-connect is the packaged onboarding skill. First use verifies the current account via get_profile, resolves the organization separately, and performs one bounded read through Code Mode. Ordinary remote workflows need no CLI. Account identity, organization selection and capability authorization are separate facts. The profile is authenticated, read-only and independent of organization/scopes; Code Mode remains the action and composition surface.

Run the scenarios in validation.json on a fresh installed chat. Record the host, package hash, actual account/org, returned receipts and pass/fail per case. Package tests do not prove OAuth, hook delivery or host behavior. Local hooks require a local execution host and trust; cloud onboarding uses native tools. Validate Claude Code separately when that host is in the release scope.

This directory is the shared source for Tedix skills used by ChatGPT Work, Codex, and Claude Code. Each skill covers one bounded task. The root portable manifest packages them for ChatGPT Work and Codex. Claude Code uses .claude-plugin/plugin.json. Both hosts load the same .mcp.json connection and skill directories. Host-specific settings must not copy the procedures or create another credential, skill, or approval store.

Use these requests to check host discovery after packaging:

On a local Codex or Claude Code host, use the installed tedix CLI and run tedix auth status before a live call. In a host without a local shell, use its authenticated Tedix MCP connection. The owner completes browser sign-in and consent; agents never request or store a copied token. Select a workspace per command with -w <workspace> instead of changing the saved default.

The canonical Work execution procedure is the Tedix organization skill work-item-operating-protocol; the public agent guide explains first use. The session guide chooses a useful Tedix path for a relevant request. It reads bounded Work state before suggesting execution, asks for a selection when more than one item fits, and keeps unrelated chats quiet. It does not create Work, record a conversation, or grant authority by loading a skill. The host skill format follows the OpenAI and Claude Code guides.

Use Tedix without the CLI

Connect the plugin through the host's OAuth flow, select your organizations, and start with the task you want done. The native connection and bundled skills support Work reads, scoped tools, tedi delegation, and Workspace Outputs without a local CLI, subject to the actual grant and organization policy. The repository connection offers the shared human Connect permission catalog, with reads selected initially in browser consent. Writes and destructive actions require explicit selection. Review the scopes in the host consent screen; .mcp.json declares only the server type and URL. connections.read permits reviewed reads through verified provider connections. Provider writes require explicit connections.execute consent; destructive actions require connections.admin. Platform writes still need their relevant consent.

tedix-session-guide chooses the useful path. tedix-delegate submits an explicitly requested task to Home and follows its durable run receipt. Neither skill silently creates Work, starts a local Attempt, or records unrelated chats. The CLI adds repository context recovery, shell automation and governed local execution. A repository policy can require that path.

ChatGPT Work and Codex installation

See the ChatGPT Work and Codex first-use guide for host-specific installation, consent, verification, and troubleshooting.

With an installed Tedix CLI, run tedix setup agents to detect Codex and Claude Code, install this plugin through each host's marketplace, and verify the installed entries. The CLI does not authorize either host's MCP connection or trust a Codex hook; complete those steps in the host.

tedix setup agents --status reports the plugin and remaining host steps. tedix setup agents --update --dry-run previews a refresh; run without --dry-run to apply it. Git marketplaces refresh through their host manager. For a local marketplace, update its checkout first: setup preserves that source.

Without the CLI, install from the generated public plugin repository: claude plugin marketplace add tedix-hq/tedix-plugins then claude plugin install tedix@tedix-plugins --scope user, or codex plugin marketplace add tedix-hq/tedix-plugins then codex plugin add tedix@tedix-plugins. It ships skills and the MCP connection without hooks.

For development, the repo marketplace at .agents/plugins/marketplace.json offers the tedix plugin. Add this checkout as a local marketplace with codex plugin marketplace add <repo-root>, then install tedix@tedix-repo from the Plugins Directory or codex plugin add tedix@tedix-repo. For a Codex CLI read-only check, first enable the MCP protocol feature Tedix Connect requires (codex features enable mcp_2026_07_28 when codex features list shows it disabled, then restart Codex), run codex mcp login tedix --scopes mcp:work.read and complete the owner OAuth flow. In ChatGPT Work, enable the local marketplace in the desktop app and start a new Work chat with the plugin enabled. Register or authorize the bundled remote Tedix MCP connection through the host's OAuth flow; the plugin contains only the gateway URL, never a bearer token. A web installation cannot install the Tedix CLI the local hooks run. ChatGPT on the web is not supported yet: the plugin is not in the ChatGPT plugin directory, and a developer-mode custom connector is not a tested path.

The plugin's lifecycle hooks are declared in hooks/hooks.json, the hosts' default plugin hook location. Review them in the host's /hooks view; Codex runs them only after you trust them. Every hook runs tedix hooks <name> behind a guard: without the tedix command it exits 0 with no output, so a host without the CLI sees no error and no live authentication or Work facts are read. With the CLI, each hook still does nothing until you opt in to its feature. The CLI guide lists every hook, when it runs, what it sends and how to turn it off; disabling or uninstalling the plugin removes them all.

The SessionStart hook stays silent unless a repository binding or TEDIX_PLUGIN_PREFLIGHT=1 opts in. Set TEDIX_WORKSPACE to choose a CLI profile and TEDIX_WORK_ITEM_ID to include one read-only Work Item summary. It reads auth and board context through the installed tedix CLI at startup, resume, clear, and compaction, so the context is restored after a host context reset. It does not log turns, start Attempts, or change Work state. With TEDIX_WORK_ITEM_ID, it also shows the latest bounded Attempt state and lease expiry, which still must be verified before a write. Command hooks cannot invoke the agent's native MCP tool session through the documented host interface, and ordinary Chat does not run local hooks.

The source hooks/hooks.json registers both await-reply (Claude Code) and await-draft (Codex); a marketplace install from this directory gets both, and each exits at once on the other host. Packaged --host artifacts ship only their host's hook.

The opt-in turn-status reporter is described in the Claude Code guide. It is silent until ~/.tedix/agent-status.json enables it.

Invoke tedix-guardian-session for an explicit start, checkpoint, or close review. It reads current CLI and repo evidence and reports state, evidence, and next action. The preflight only points to this routine when opted in; it does not run in the background.