TAG40@@Everyone on the team can now use coding agents. What still feels off is that those agents are still inside each person's terminal. Decisions are scattered across chat histories, and incorrect "facts" may be treated as true by the next agent to pick up the work.

In this article, we explain how to connect Claude Code and Codex to Ontoloop via a remote MCP and perform the following tasks within the same team workspace.

• Read typed team facts that carry a version with every change
• Propose writes that will not be applied until a person verifies them
• Accept tasks handed over by a colleague and report progress in a thread

Ontoloop is currently available through invite-only early access. To request access, email hello@ontoloop.ai. The web app's UI is currently in Chinese. The commands and tool names below are based on the published MCP documentation. This article is a technical explainer by Ontoloop's official account.

What You'll Need

• An Ontoloop account (invite-only early access; request access from hello@ontoloop.ai)
• A remote MCP client that supports Streamable HTTP and browser-based OAuth
  • Claude Code
  • Codex
  • or Claude desktop / Claude on the web / WorkBuddy (a connector for pasting a URL)
• The remote endpoint (according to the MCP documentation (updated 2026-11-04), only the URL goes into the setup command)

According to the documentation, the transport is Streamable HTTP, and authentication uses OAuth. When the client first connects, a browser opens, and you click Allow (允许) on Ontoloop's authorization page. The setup command contains only the URL, and authorization is performed in the browser.For clients that cannot sign in via a remote machine, SSH, or a browser, the documentation guides you through manual setup on the My Agent (My agents) page.

1. Connect Claude Code

To make it available across projects, add the connector at the user scope.

Next, perform the following steps in order within Claude Code.

1. Enter /mcp
2. Select ontoloop
3. Select Authenticate
4. On the Ontoloop authorization page that opens in your browser, click Allow

The -s user option is required because Ontoloop is intended to be used across projects rather than within a single local folder.

If the authorization expires, use the following command.

2. Connect Codex

The authorization page opens in your browser, so click Allow. If it does not open, run the following.

3. Other clients (URL only)

Client

How to connect
Claude desktop / web	In Customize → Connectors → Add custom connector, paste https://ontoloop.ai/api/mcp, then click Add and Allow
WorkBuddy	In Settings → Connectors → Add, enter only https://ontoloop.ai/api/mcp, then click Allow

The following is additional information from the documentation.

• Claude's free plan supports only one custom connector.
• With the Team and Enterprise plans, the organization owner first adds a connector in the organization settings, and then members click Connect.
• For clients on remote machines, for SSH connections, or for clients that cannot sign in through a browser, use the manual setup on Ontoloop's My agents page. This article does not describe the steps that follow.

4. You get an Agent after connecting

The Agent you connect to is assigned a unique identity under your name, such as “Lawrence's Codex.” Its capabilities do not exceed your own, and you can disconnect or revoke it from My agents.

The official list of tools is what tools/list returns on the client. The MCP documentation includes the following commonly used groups. Although the public README and the MCP registry's server.json specify version 0.1.0, this article does not verify tools/list on a live server.

Purpose	Tools
Read the team's facts	list_spaces, list_types, search, list_objects, read_object, read_artifact
Propose and reference	propose, reference, upload_artifact, list_my_proposals, get_proposal
Receive work from colleagues	list_my_tasks, accept_task, decline_task, update_task_progress, submit_task_result
Work together in threads	start_thread, list_threads, read_thread, post_update

The public MCP README also lists the Knowledge, Extraction, Skills, Apps, and Guide groups (for the latest complete list, see tools/list). This article focuses on the four groups above, covering shared facts, person verification, tasks, and thread workflows. The website lists Org-wide Skills as “Coming soon,” so this article does not treat that group as currently available to everyone.

Trying an Agent: a practical loop

The following are example natural-language prompts, not validated API payloads. They refer only to the capabilities of the tool groups described in the documentation. You can try them once authorization succeeds.

1. First, get the lay of the land: “List the spaces I can access and show the object type of the current space.”
2. Read before writing: “Search for already confirmed decisions about <topic> and summarize them.”
3. Do not invent facts without verification: “Propose an update to that decision with a brief rationale. Do not treat it as final until it is confirmed.”
4. Check status: “List my unconfirmed proposals and show the rejection reason, if there is one.”
5. Tasks: “List my tasks, accept one, write a progress update in the task’s thread, and submit the result.”
6. Threads: “Start a thread to summarize what changed and what still requires human confirmation.”

The exact argument schemas are what the client displays from tools/list. This article does not provide JSON payloads or extra CLI flags.

5. Why writes require human confirmation

This is not a UI convenience, but a central product rule.

• Decisions, tasks, and other facts written by an Agent first enter Pending (awaiting confirmation) and do not take effect yet.
• The Agent’s owner or a space administrator can review, edit and then confirm, or reject them on the web.
• Only confirmed items become team facts that other Agents can read and cite.
• Uploads of regular artifacts (files) skip confirmation and are stored as versions. Uploading the same content again does not add another version.
• The exception is a new version of an existing Skill, as stated in the MCP documentation. The upload remains pending and is not applied until the owner confirms it on the web. Since the homepage labels organization-wide Skills as “Coming soon,” check what is available in your own workspace before relying on this procedure.
• The space’s activity log records who, which Agent, what was changed, and when.

This prevents all Agents from treating unreviewed speculation as shared truth.

6. Tasks and threads across colleagues

Once facts are shared, you can do the following.

• Tasks assigned by colleagues appear in list_my_tasks.
• The Agent can accept_task or decline_task, update_task_progress, and submit_task_result.
• Interactions with the owner are carried out in threads (start_thread, list_threads, read_thread, post_update). Markdown updates can also be used.

When one person investigates in Claude Code and another continues in Codex, both Agents should cite the same verified object rather than reconstructing decisions from their respective chat histories.

7. FAQ from public documentation

Does the setup command include a token?
According to the MCP documentation, no. Only the server URL is included. Authorization is performed in the browser. For remote servers, SSH, or clients that cannot sign in through a browser, the documentation directs users to the manual setup on the My agents page.

What happens when authorization expires?
For Claude Code, use claude mcp login ontoloop. For Codex, use codex mcp login ontoloop (in the desktop app, click Authenticate in the MCP list under Settings). For Claude desktop and web, open Ontoloop in the connector list and click Connect. For WorkBuddy, reconnect Ontoloop from Settings → Connectors.

Where is the server’s source code?
Ontoloop is a hosted service. The public repository contains a description of the remote MCP server and the server.json used by the MCP directory. The registry name is ai.ontoloop/ontoloop.

Links

• Product: https://ontoloop.ai/en
• MCP documentation: https://ontoloop.ai/en/docs/mcp
• Public MCP repository / README: https://github.com/alextangson/ontoloop-mcp
• Early access: hello@ontoloop.ai