# Amp MCP: Give Your Agent Real Bug Evidence _2026-11-05 • 9 min read • By clip.qa team_ **Category:** Engineering **Keywords:** Amp MCP; agentic qa; llm bug reporting; describe bugs to ai concisely --- ## Overview Amp MCP connects a coding agent to structured bug-report evidence instead of forcing the agent to infer behavior from a vague description. clip.qa exposes a remote Streamable HTTP MCP endpoint at https://mcp.clip.qa/mcp , protected with OAuth2.1 PKCE. The agent can list reports, inspect details, retrieve network requests or console logs where available, compare two reports, and reference a fix or fix attempt. --- ## Introduction

Amp MCP connects a coding agent to structured bug-report evidence instead of forcing the agent to infer behavior from a vague description. clip.qa exposes a remote Streamable HTTP MCP endpoint at https://mcp.clip.qa/mcp, protected with OAuth2.1 PKCE. The agent can list reports, inspect details, retrieve network requests or console logs where available, compare two reports, and reference a fix or fix attempt.

## What does Amp MCP provide to a coding agent? The important boundary is task data. A report can contain text, screenshots, video, and recorded interaction evidence, but it should remain evidence to evaluate—not instructions for the agent to execute. Pair every claimed fix with a retest in the original environment. Amp runs as a coding agent in its development environment and can use code, command-line tools, and remote capabilities. Its documentation also distinguishes local MCP definitions from remotely managed definitions, which matters when the same repository is used from more than one machine or execution context. Amp documentation confirms the agent environment and its support for orbs—machines per thread with code and tools. With clip.qa connected, the agent can work from a report ID rather than reconstructing a defect from memory. Relevant remote tool names include listReports, getDetails, getNetworkRequests, getConsoleLogs, getScreenshots, getVideoTranscript, getReportDiff, createComment, linkFix, and linkFixAttempt. This supports agentic QA because the agent can begin with a narrow question, gather relevant evidence, change code, and request a comparison. It does not mean clip.qa changes the code, guarantees a correct diagnosis, or validates every fix by itself. For a concise evidence brief, use the pattern in how to describe a bug to AI. Include observed behavior, expected behavior, exact reproduction steps, platform, and a stable report or task reference. Avoid telling the agent that “the app feels broken.” ## How do you add clip.qa to Amp with OAuth? Amp supports remote OAuth through its managed remote-definition workflow. This is a documentation-based setup, not an end-to-end integration test. Open a terminal in the environment where Amp runs. Add the personal remote definition and request explicit scope: amp mcp remote --personal add clip-qa https://mcp.clip.qa/mcp --auth oauth Complete the browser authorization flow when prompted: amp mcp remote --personal login clip-qa Check the server and authentication state: amp mcp remote --personal check clip-qa Refresh the exposed tool list: amp mcp remote --personal tools clip-qa --refresh Start a fresh Amp thread if the new server or tools do not appear in the current session. Use the personal scope shown here unless your own access policy requires another arrangement. Amp’s MCP documentation explains that managed remote definitions and local configuration are separate concerns. A local URL in amp.mcpServers is not automatically synchronized into a remotely stored definition. The clip.qa server uses OAuth2.1 PKCE. Do not paste access tokens, client secrets, or credentials into repository files. Keep write-capable operations, such as createComment, linkFix, and linkFixAttempt, separate from the initial read-only investigation whenever your approval model allows that distinction. ## Which Amp execution context can see the MCP server? Start by identifying where Amp is running. A local editor process, a terminal process, and a remote orb may not load the same MCP definitions. If check succeeds locally but an orb cannot list tools, inspect the remote environment rather than repeatedly reauthorizing locally. Use this troubleshooting sequence: Run amp mcp remote --personal check clip-qa in the failing context. Refresh tools with amp mcp remote --personal tools clip-qa --refresh. Ask the agent to list reports through listReports; do not begin with a write tool. Open one known report through getDetails and confirm the identifier matches. If local access works but remote access fails, compare how the local and managed remote definitions were created. Record the failing context, command output, and tool availability before changing repository code. Do not assume that a local amp.mcpServers entry becomes a managed remote definition. Amp documents remote definitions as a separate mechanism. Also do not assume every cloud execution surface inherits local configuration. ## How should an agent investigate a pagination regression? clip.qa’s iOS and Android apps and Chrome extension are in beta; request access before planning a capture. Video analysis runs in the cloud and produces a draft for a person to review. Console and network capture applies to the browser extension, not arbitrary native mobile apps. This setup follows the published documentation; it has not been tested end to end here. See the clip.qa MCP page for product details. Consider a hypothetical report: a product list initially shows 20 rows. After changing a date filter, the interface still shows 20 rows, but the returned page belongs to the previous filter. The screen recording shows the selected filter and a stale row count. The report also includes an automated test note stating that the request parameter changed while the client continued to display the previous cursor state. A disciplined agent should not immediately patch pagination. It should collect comparable evidence first: Call listReports and select the report by a stable identifier. Call getDetails for the same report and preserve the reproduction steps, expected result, and actual result. Read transcript narration for context, then check the recording and screenshots to verify the interaction order. Use network-request evidence only when it exists for the captured environment. Compare the filter-change event, request parameters, response, and rendering sequence. Locate the client state transition and pagination cursor handling in the repository. Write a regression test that changes the filter and verifies both the request and visible rows. Run the repository’s own baseline tests before and after the change. Repeat the original user flow in the original platform and context. Record the fix attempt with linkFixAttempt only after the retest result is known. This example is hypothetical, but the validation logic is practical: code tests and the original environment answer different questions. An automated test can prove a narrow state transition. An original-flow retest can confirm that the user-facing flow no longer shows stale results. ## How do you compare report evidence without trusting stale results? Pagination problems often produce contradictory snapshots. A report may show the filter before it changes, while a later screenshot shows the changed filter with old data. Use getReportDiff to compare two reports, but treat the diff as a comparison aid—not a complete causal explanation. Ask the agent to label each item by source: screen evidence, written reproduction notes, network evidence, console evidence, transcript, or code inspection. Missing browser console or network data must remain missing; do not invent it. For background on structuring evidence, see LLM-ready bug reports. If evidence conflicts, narrow the claim. For example, say: “The recording shows the changed filter, while the stored request detail shows a different parameter value.” Do not collapse that into “pagination is definitely broken.” The narrower statement gives the next investigation a precise target. ## How do you hand a fix back to reviewers? Use createComment for review context and linkFix for the code change. A linkFixAttempt log can be pending while validation is underway. Mark it succeeded or failed only when the original-environment retest supports that outcome. A reproducible handoff should include: Report or task identifier. Expected and observed behavior. Root-cause hypothesis and supporting evidence. Files and symbols changed. Automated checks run and their actual results. Original-environment retest steps and result. Remaining limitations, including unavailable device telemetry. For mobile defects, remember that browser capture tools do not automatically expose arbitrary native app console or network data. iOS telemetry can also be limited while the target app is in the foreground. Keep those limits visible in the handoff rather than implying complete instrumentation. Reports remain private until explicitly shared, so verify recipients before posting sensitive context. Sign up free to capture evidence for an AI-assisted debugging workflow. ## Sources Amp documentation — Confirms the coding-agent environment and orbs with code and tools. Checked 2026-11-05. Amp MCP documentation — Confirms local versus managed remote definitions and the documented remote OAuth workflow. Checked 2026-11-05. clip.qa MCP documentation — Confirms the Streamable HTTP endpoint, OAuth2.1 PKCE authentication, and remote tool names. Checked 2026-11-05. ## Key Takeaways - The important boundary is task data. - Amp supports remote OAuth through its managed remote-definition workflow. - Start by identifying where Amp is running. - clip.qa’s iOS and Android apps and Chrome extension are in beta; request access before planning a capture. ## Frequently Asked Questions **Q: Does adding Amp MCP install code into the tested app?** No. This workflow connects the coding agent to clip.qa’s remote MCP server; it does not install an SDK into the application being tested. **Q: Which clip.qa tool should an agent call first?** Start with listReports , then call getDetails for the selected report. This keeps discovery and detail retrieval explicit before using screenshots, transcript, network, console, or diff tools. **Q: Can an agent use clip.qa to mark a fix successful?** An agent can create a fix-attempt record with a commit or pull request and a pending, succeeded, or failed outcome, but success should reflect a real retest in the original environment rather than an automated test alone. **Q: Why does a local Amp MCP setup not appear in a remote thread?** Local and managed remote MCP definitions can be separate in Amp. Check the remote definition, authentication, and tool visibility instead of assuming a local URL is synchronized. **Q: Is an authenticated clip.qa MCP server available on every plan?** Yes. clip.qa MCP is available across all tiers, including Free, subject to your account and server authorization state. ## Related Posts - [Read more](/blog/how-to-describe-bug-to-ai) - [Read more](/blog/llm-ready-bug-reports)