> ## Documentation Index
> Fetch the complete documentation index at: https://docs.feedbackloop.page/llms.txt
> Use this file to discover all available pages before exploring further.

# Coding agents

> Connect selected Feedback Loop projects to Codex, Claude Code and other MCP clients.

Use your coding agent to review requests, implement selected improvements and prepare customer updates with the context from your Feedback Loop project.

## Connect your agent

Add a **Streamable HTTP** MCP server in an OAuth-capable client with this address:

```text theme={null}
https://feedbackloop.page/api/mcp
```

Start the client's authentication flow, sign in to your Feedback Loop owner account and review the client name. Select only the projects it needs and choose access for each project:

| Access | What the agent can do |
| - | - |
| **Read** | Read the selected project's feedback, discussions and release notes, including unpublished drafts. |
| **Manage** | Everything in Read, plus change feedback statuses, post public comments, create or edit release drafts and publish releases. |

Projects outside your selection remain unavailable. Access is checked separately for each project, so one connection can have Read access to one project and Manage access to another. The connection uses your owner authority; community website sign-in keys are not MCP credentials.

> [!WARNING]
>
> Read access includes private release drafts. Manage access allows public
> changes. Grant access only to a client you trust and tell the agent which
> actions you want it to take.

## Download the plugin

[Download the Feedback Loop plugin](https://feedbackloop.page/integrations/feedback-loop.zip) and extract the archive. It contains one `feedback-loop` folder with the same five workflows for Codex and Claude Code. The download is hosted by Feedback Loop; you do not need access to a source repository or a marketplace listing.

For Claude Code, run this from the folder containing the extracted plugin:

```sh theme={null}
claude --plugin-dir ./feedback-loop
```

Then open `/mcp` to authenticate its `feedback_loop` server. For Codex, register the extracted folder in your approved local or team marketplace using the [official plugin packaging guide](https://developers.openai.com/plugins/build/plugins).

You can also connect MCP tools directly without downloading the bundled workflows.

For a direct Codex CLI connection:

```sh theme={null}
codex mcp add feedback_loop --url https://feedbackloop.page/api/mcp
codex mcp login feedback_loop --scopes mcp:read,mcp:manage,offline_access
```

For a direct Claude Code connection:

```sh theme={null}
claude mcp add --transport http feedback_loop https://feedbackloop.page/api/mcp
```

Open `/mcp` to authenticate the configured server. Use the client's normal OAuth flow rather than copying tokens or session cookies into a prompt. Configure either the plugin's server or a direct connection to avoid loading the same tools twice.

## Choose a workflow

* **Review feedback:** ask for themes, supporting requests and priorities. The review leaves project content unchanged.
* **Implement feedback:** ask for a selected request to be implemented in your repository. The agent follows that repository's instructions and verification restrictions. Implementation does not automatically change the request's status.
* **Draft release notes:** prepare customer-facing copy. With Read, the agent returns copy in chat. With Manage and a request to save, it creates or updates an unpublished draft.
* **Reply to feedback:** ask for a draft response, or explicitly ask the agent to post to a selected discussion. Posting requires Manage.
* **Close the loop:** name the release, feedback items and actions you want, such as publishing a reviewed note or changing a confirmed deployed request to Shipped. Publication and public replies require explicit instructions.

For example: “Review the latest feedback for My Project and link the requests behind each recommendation.” Or: “Draft release notes for these changes and leave them unpublished.”

Ask for source links and the range reviewed when prioritizing feedback. Lists are paginated in groups of 25; a partial review may miss older requests or replies. Passing tests or merging code does not establish that a feature is deployed.

## Review or disconnect access

Open account **Settings → Connections** to see each authorized client and its selected projects and access levels. Choose **Disconnect** to revoke a connection. To change its project selection or access, disconnect and reconnect through the client.

Disconnecting one connection does not disconnect other independently approved connections. If the client reports lost access, check that its connection still exists and that the intended project is selected before reconnecting.

If a write times out, check the discussion or release list before asking the agent to try again. A response can be lost after the content was saved. The bundled workflows investigate uncertain writes and report partial completion instead of automatically creating duplicates.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.