Skip to content

Connect your agent

Choose the setup that matches where you use your agent. Both use your existing Tylor account.

The public Tylor landing page introduces ChatGPT access under ChatGPT. Choose Set up ChatGPT there to return to these instructions. Your administrator must enable access; signing in to Tylor alone does not enable it.

Once access is enabled, connect on ChatGPT web:

  1. Open Settings → Security and login and enable Developer mode.
  2. Open Plugins, choose +, and name the connection Tylor.
  3. Set the MCP server URL to https://connect.tylor.app/mcp and choose OAuth. If offered a client registration choice, choose CIMD. No API key or client secret is needed.
  4. Sign in with your existing Tylor account, select the folders ChatGPT may access, review the permissions, and choose Connect Tylor. Select at least one folder; none is selected for you. A sample-only request instead shows Connect pilot. Decline cancels either request.
  5. Install the plugin, start a new Work conversation, and select Tylor using @.

The available settings depend on your account and workspace. These steps follow OpenAI’s plugin quickstart.

See the ChatGPT connection formChatGPT New Plugin form named Tylor, with https://connect.tylor.app/mcp as the server URL and OAuth selected.
See the folder consent screenTylor consent with Research selected, a searchable folder list, document and package permissions, and Connect Tylor and Decline buttons.

Choose up to 20 folders. ChatGPT can only act within that selection and your current Tylor permissions.

Where What ChatGPT can do
Selected folders and their subfolders Access current documents, plus documents added or moved in later.
Workspace root or outside the selection No access. Move the document into a selected folder first.
Shared folders Read or edit according to your role. Viewers cannot edit.
Folders you own within the selection Create new documents.
  • Folder changes: renaming keeps access; moving outside the selection, deleting, or losing access removes it. An explicitly selected subfolder keeps its own selection if moved.
  • No folders yet? Create one in Tylor, then choose Reload folders.
  • Older connection? Reconnect to select folders and approve package access. If old permissions persist, remove it and connect again. Sample-only connections never upgrade automatically.

Edits save immediately—there is no tracked-change acceptance step in Tylor.

  • Recover: open History for the previous content.
  • Avoid conflicts: ChatGPT reads before editing. Conflicting saved edits or pending suggestions stop replacement; resolve suggestions in Tylor.
  • Sync first: unsynced browser edits are not visible to ChatGPT. Saved changes update collaborators and any existing public PDF.

Ask: “In Connection test, add a house-style package with a reusable title and page style, then update main.typ to use it.”

Task How it works
Use an existing style ChatGPT reads the document’s local packages and referenced dependencies.
Add a style Approve Add new styling package versions. You must own the document.
Revise a style Add a new version and update the import. Existing versions cannot be overwritten or deleted.

Packages stay in the document. My library is unchanged. The previous import is recoverable through History.

To find them, open the document, then choose Account → Local packages…. Scroll below the publishing form to This project and choose Open to edit. Opening Local packages from the dashboard shows only your account library.

For example, the document can import its style with #import "@local/house-style:1.0.0": style.

File limits and unsupported operations
Limit Maximum
Each text file 256 KiB
New document or package version 20 text files, 512 KiB total
  • Local package dependencies must already exist. Unrelated library entries and package-source projects are not exposed.
  • Use Local packages for binary assets and account-wide package management.
  • Structural file edits, package copying and real-document PDF export are not available through ChatGPT. Download PDFs from Tylor.
  • Large projects may exceed the History snapshot limit; edit them in Tylor.
  • After a timeout, the agent should retry the same request ID and content to avoid duplicate writes.

With Tylor selected in your conversation, try these requests in order:

  1. “List my Tylor projects.” Check that you recognize the projects returned.
  2. “List the folders Tylor allows you to use, then create Connection test in [folder name].” Choose a selected folder you own and open the resulting document link in Tylor.
  3. “Read its main.typ and add a short introduction.” Check the saved result and its previous version in History.
See Tylor selected in a ChatGPT conversationChatGPT conversation composer with Tylor selected above the message field.

Repeat from ChatGPT on your iPhone if that is where you plan to use Tylor.

Optional: check the sample tools and PDF link

Ask “List the sample projects.” The expected result is Synthetic MCP test, ID sample-project, identified as synthetic data. OAuth linking and this tool call have been confirmed by a tester. If the connection screen says No app actions available yet, choose Refresh and try the request in a new conversation; that display alone does not prove the tools are unavailable.

Next ask “Return the sample PDF as a link.” Open the link and repeat from ChatGPT on your iPhone. A tester confirmed the sample PDF opens on a physical iPhone. This verifies link delivery; it does not guarantee an in-chat attachment. After at least ten minutes, ask for the sample projects again to check whether ChatGPT renews the connection without another login.

Open Account → Connected apps or Connected apps to review permissions and choose Disconnect. The page shows the folders selected at consent; reconnect from ChatGPT to change the selection. A request saying it expired or was already used needs a fresh connection from ChatGPT. If ChatGPT access is disabled, ask your administrator to enable it. After a failed refresh, reconnect; reusing an old refresh credential revokes that connection. New requests stop after disconnect; an already-authorized write may finish. Previously returned text and downloaded files remain in the receiving app.

Use this setup for Codex, Claude Code, Pi, OpenCode, or another agent that can run terminal commands and read skill instructions. You need a signed-in Tylor account and a current Node.js LTS version in the environment where the agent runs (CLI 0.6.0 requires Node 18 or newer).

Install the CLI, add the skill, sign in, then verify the connection.

Terminal window
npm install -g tylor-cli@latest
tylor --help

The package is named tylor-cli; its command is tylor. Use the current CLI with a current Tylor backend, especially when an agent and a person edit the same project.

Installation failed? See CLI troubleshooting for macOS permissions errors, older Node versions, and commands that cannot be found.

Choose the agent you use. These commands install the same tylor/SKILL.md into its local, user-level skills directory. Paths below use macOS/Linux shell syntax; use the corresponding home-directory path in your environment.

Install for CodexIllustrative session
tylor skill install ~/.agents/skills
installed the Tylor skill → /home/you/.agents/skills/tylor/SKILL.md

Sample output. Commands are not executed on this page.

Codex skill locations include ~/.agents/skills. If the skill does not appear, restart Codex. Older Tylor instructions used ~/.codex/skills; use the location supported by your installation and avoid duplicate enabled copies.

The previews show an excerpt of the installer output. Install only where needed, and refresh or restart the agent if the skill is not detected.

Run the browser-assisted setup (CLI 0.4.0 or newer):

Terminal window
tylor login

Tylor opens in your browser and reuses your sign-in. Check the account and matching terminal code, select I started this login and the code matches my terminal, then choose Connect CLI. The CLI creates and saves its own API key; you do not need to copy a secret or edit a file. The terminal confirms when the login is saved.

The key acts as your account, including reading, changing and deleting accessible data. Only approve a request you started. You can revoke the device key in Account → Advanced → API keys.

For a remote terminal or a browser that does not open, run tylor login --no-browser in the agent’s execution environment and open the printed link on your own computer. Verify the code before approving. Installing on your laptop does not configure a container or remote agent.

Login saves ~/.config/tylor/config.json for the current operating-system user, preserving existing settings such as library backups. This also works for desktop agents that do not inherit terminal environment variables. If TYLOR_API_KEY is already set, it overrides the file: remove that override before browser login, or keep using it for automation.

See CLI and API keys for screenshots, self-hosted instances, reconnecting an account, and manual key setup. See CLI troubleshooting if the browser, connection, or config save fails.

With CLI 0.6.0+, run tylor login --alias company to add a second account and approve that identity in the browser. Existing logins remain saved. Use tylor auth list to see them. Then tell your agent: “Use company@example.com for this report.” The updated skill verifies the account and includes --account on every command for that task. You can use a different account in another conversation without switching the global default.

See multiple accounts for aliases, defaults, environment overrides and upgrading old checkouts.

First run tylor list in your terminal. Then ask your agent:

Use the Tylor skill to list my projects. Do not change anything.

Confirm that it finds the intended workspace. You can then give it a project name or ID and a task. See work with your agent for a terminal walkthrough.

Run npm install -g tylor-cli@latest, then repeat the skill-install command for your agent above. The installer replaces that copy of the Tylor skill. Keep personal workflow instructions in a separate skill so an update does not overwrite them.

Symptom Check
tylor not found Node and the global npm executable directory must be available to the agent, not only your interactive terminal
Skill not found Check the installed tylor/SKILL.md path and refresh or restart the agent
No API key Run tylor login in the same machine/user environment where the command runs
Unauthorized or forbidden Check the key, account status, project access, and API endpoint
Network or permission error Allow the required CLI execution and API access in your agent environment
A project is missing Check the account; browser-only local projects are not available through the API