Crash reports for developer tools, generated by AI agents.
AI agents are the new primary consumers of APIs, SDKs, and developer documentation. When an agent fails to make a tool work, the failure is silent — the developer moves on, the vendor never learns what went wrong.
Before automated crash reporting, app vendors had no visibility into user-side failures. The same gap exists today for developer tools: vendors have zero signal when an agent struggles with their API, gives up on their auth flow, or abandons an integration because a needed capability doesn't exist.
A standardized way for AI agents to generate structured feedback about their experience with developer tools, and — with the developer's explicit approval — deliver it to the tool vendor.
The developer controls everything. Reports start fully redacted. The developer reviews the draft, decides what to include, and explicitly approves before anything is sent. This isn't telemetry. It's a crash report the developer writes with AI assistance.
sep/ # MCP Specification Enhancement Proposal
agent-implementation-feedback.md
skill/ # Claude Code /feedback skill
feedback.md
schema/ # Feedback report JSON schema
feedback-report.schema.json
examples/ # Real-world example reports
grain-mcp-feedback.json # MCP auth breaking across versions, no batch endpoint
instantly-api-feedback.json # Undocumented ESP 999 errors, no rate limit headers
linkedin-enrichment-feedback.json # Search timeouts, needed 2-tier fallback strategy
beehiiv-api-feedback.json # Undocumented rate limits, generic error responses
prismic-migration-api-feedback.json # UID caching after deletion, silent 404s after publish
vercel-nextjs-caching-feedback.json # Edge cache not clearing, breaking API changes
- You're using Claude Code (or Cursor, or any AI agent) to integrate a tool
- The integration fails or is impractical
- You invoke
/feedback - The agent drafts a structured report based on the session context
- You review every section — remove, edit, or redact anything you want
- You approve and send (or copy/save for manual submission)
- Declare a feedback endpoint in your MCP server capabilities (or accept reports via GitHub issues)
- Receive structured reports from agent platforms
- Aggregate reports to identify friction patterns
- Fix the issues that are causing silent abandonment
Reports include:
- Tool identity — what tool, what version, what protocol
- Session outcome — completed, completed with workarounds, or abandoned
- Friction points — categorized (auth, missing capability, error clarity, docs, etc.) with severity and workarounds attempted
- Narrative — AI-summarized implementation story (not raw logs)
- Suggestions — actionable improvements
Reports never include: source code, file paths, credentials, conversation transcripts, or environment variables.
The sep/ directory contains a draft Specification Enhancement Proposal for the Model Context Protocol. It proposes adding a feedback capability to MCP so that agent platforms can deliver implementation feedback to MCP servers through the protocol itself.
This started with a failed MCP integration. We tried to pull 67 meeting transcripts from Grain's MCP server in Claude Code. Auth broke across sessions due to transport version mismatches. There was no batch endpoint, making the bulk workflow impractical. We abandoned after 3 sessions.
Claude Code had full context of every failure — every config attempt, every workaround, every dead end. But to tell Grain, we had to manually draft a message and submit it through a web form. That feedback loop is broken.
See examples/grain-mcp-feedback.json for what the automated report would have looked like.
The examples/ directory contains feedback reports from real integrations built with Claude Code across two projects — a GTM automation platform and a Next.js website with Prismic CMS:
| Example | Tool | Protocol | Outcome | Key Friction |
|---|---|---|---|---|
grain-mcp-feedback.json |
Grain | MCP | Abandoned | Auth broke on version bumps, no batch endpoint |
instantly-api-feedback.json |
Instantly.ai | REST | Completed w/ workarounds | Undocumented "ESP 999" errors, no rate limit docs |
linkedin-enrichment-feedback.json |
DuckDuckGo/LinkedIn | Web scraping | Completed w/ workarounds | Search timeouts, 1.5% failure rate on valid profiles |
beehiiv-api-feedback.json |
Beehiiv | REST | Completed w/ workarounds | Undocumented rate limits, generic error responses |
prismic-migration-api-feedback.json |
Prismic | REST | Completed w/ workarounds | UID caching after deletion, silent 404s, unpushed slice failures |
vercel-nextjs-caching-feedback.json |
Vercel / Next.js | SDK | Completed w/ workarounds | Edge cache ignores revalidateTag(), breaking API change in v16 |
Six tools. Six sets of friction. In every case the agent had full context of the failure but the vendor received zero signal.
This is early-stage. If you've hit similar walls — agent struggled with a tool, vendor had no idea — we'd like to hear about it. Open an issue with your experience.
If you're a tool vendor interested in receiving agent feedback, open an issue and we'll work with you on the endpoint spec.
MIT