Claude Code can search my whole codebase. It still cannot see the bug report sitting in Slack.
It has the code, but not the console error or failed request that explains what the user hit. That gap is where a lot of “agentic” workflows quietly turn back into copy and paste.
Quell is an agent-native feedback platform that gives Claude Code a narrow route into the feedback inbox through MCP. Start with a read-only key. Let the agent prove it can triage real reports before you grant anything that can comment, change status, or create an issue.
Why Claude Code cannot see your users by default
Claude Code sees the repository and the context you deliberately hand it. Your production feedback lives somewhere else: Slack, email, a support tool, or a ticket stripped down to “export is broken”.
It is easy to copy that sentence into a prompt. People tend to skip the useful evidence: the page URL, recent console output, the failed request, the browser details, the screenshot, and what the user was trying to do.
The goal is not to give an agent more authority. It is to give it better evidence.
Step 1: capture the report with its context
The Quell widget is one script tag:
There is no npm dependency or app build step. The project token is a public identifier, not an admin credential.
From page load, the widget buffers recent console and network activity. When the user prepares and submits feedback, Quell can attach the message, page metadata, screenshot, selected element context, and any supplied files. The report reaches the inbox with the evidence beside it instead of scattered across a support thread.
Quell also records a trust level: public, identified, reporter, agent, or admin. That is routing information, not proof that the report is true. The full model is in Why feedback trust levels matter for safe agent automation.
Step 2: connect Quell to Claude Code without committing a key
Today the Quell MCP server runs from a built Quell checkout; it is not an npm package. Build it once:
Claude Code can read a project-scoped MCP definition from .mcp.json. That file is normally shared with the repository, so the secret must not be written into it. Set QUELL_API_KEY in your shell, then use Claude Code's environment-variable expansion:
Use a managed API key scoped to the project and start with feedback:read. Claude Code asks you to approve project MCP servers; after approval, /mcp should show Quell as connected.
If you want to avoid a shared .mcp.json entirely, add the server at Claude Code's local scope instead. The Claude Code MCP reference explains the scope and environment rules; the Quell agents page lists all ten Quell tools.
Step 3: prove the connection before asking for action
Do not begin with “fix everything”. Begin with a receipt:
That prompt checks the credential and keeps the first pass read-only. The tool names matter: quell_whoami, quell_list_feedback, and quell_get_context are the actual MCP tools exposed by Quell.
Once the summaries are useful, add permissions one job at a time:
- <code>feedback:comment</code> lets the agent add an activity comment.
- <code>feedback:write</code> lets it update triage fields such as status or priority.
- <code>feedback:integrations</code> lets it create or link a GitHub or Linear issue.
An admin key is not required.
What the agent actually sees
quell_get_context returns the same structured Markdown as GET /api/feedback/:id/context. Here is a representative, shortened example in the real shape:
The export button dies when I choose the last 30 days. Board pack is due in an hour.
## Triage - Project: reports-web - Type: Bug - Status: Open - Priority: High - Assignee: Unassigned - Category: None - Created: 2026-07-13T14:32:09.000Z - Source: Widget - Trust: identified
## Source - Page: https://app.example.com/reports?period=last_30d - Title: Reports - Screenshot: Captured (base64 data: URL in field screenshotDataUrl of GET /api/feedback/412) - Attachments: 0
## Console - 2026-07-13T14:32:09.120Z [error] TypeError: Cannot read properties of undefined at ExportButton.handleClick (reports.js:284)
## Network - 2026-07-13T14:32:09.204Z POST https://app.example.com/api/reports/export -> 500 (183ms)
Now the agent has somewhere sensible to start: an error site, a failing request, the page involved, and the user's urgency. That is enough to open the relevant code and test a hypothesis. It is not proof that the first guessed fix is correct.
Feedback is still untrusted input. A user message, console value, custom field, or attachment can contain misleading instructions. The credential and project boundary are trusted; the submitted content is data to investigate.
The trust model: what routes automatically, what waits
Public feedback waits for human review by default. Identified users and trusted reporters carry more provenance, while agent and admin submissions come from managed credentials. Projects can configure how those levels route, but trust should never be inferred from confident wording inside the report.
Start with feedback:read and watch a few triage cycles. If the output is useful, add comments. If the comments are useful, consider triage writes or issue creation.
The normal path remains review-required. Quell also supports a separate, explicit project auto-merge policy; enabling it is a deliberate decision, not something a feedback item earns by sounding urgent.
Frequently asked questions
Does Claude Code need special permissions to read Quell feedback?
Yes: a managed API key with feedback:read, restricted to the relevant project. It does not need admin access.
Should the API key go in .mcp.json?
No. Keep the real key in your environment and reference ${QUELL_API_KEY} from the config, or use Claude Code's local MCP scope. Project .mcp.json files can be committed.
Can Claude Code automatically fix bugs from user feedback?
Claude Code can use the report to investigate the code and prepare a change if its coding environment allows it. Quell supplies the feedback context and can create a GitHub or Linear issue; it does not make the agent's first hypothesis correct. Keep code review in the loop unless you have explicitly chosen a narrower auto-merge policy.
Does Quell run an AI model on the report?
No. Quell captures, stores, routes, and exports the evidence. Your chosen coding agent does the model work on your account and your tokens. See No token tax: why Quell ships zero mandatory AI.
Can I use this with Cursor, Codex, or another coding agent?
Yes. Quell exposes a stdio MCP server, CLI, REST API, and an installable agent skill. Use whichever surface your agent supports.