Vibe AI agent trends: what breaks, what's missing, what agents need
71 public reports about AI agents and MCP, found by our research workers, and a sign-in health check of 809 remote MCP servers that we run ourselves. Two kinds of data, labeled separately.
Updated 2026-11-13 01:09 UTC (refreshes hourly). Every number is a count from our database; nothing is estimated. Measured = we observed it. Reported by research workers = a research worker found and summarized a public source; it counts reports, not users or traffic. JSON for agents: /api/trends (CC BY 4.0).
Key facts
| Research findings Reported by research workers | 71 from 2 research runs (claude-research:2026-11-04: 35; claude-research:2026-11-11: 36) |
|---|---|
| Source dates | 2026-10-24 to 2026-11-11 (71 findings with a date) |
| Largest finding type | Failures: 41 of 71 |
| Found in more than one run | 0 of 71 |
| MCP servers health-checked Measured | 809 listed remote servers, latest probes 2026-11-11 16:10 UTC to 2026-11-12 07:57 UTC |
| Latest result | 442 reachable (55%), 350 need credentials (43%), 17 failed (2%) |
| Median response, reachable | 151 ms (n = 442) |
What research workers found Reported by research workers
Each finding is one public source (mostly GitHub issues) that a research run found and filed under one type. Sample: 71 findings, source dates 2026-10-24 to 2026-11-11. This shows what people write about, not how often things happen.
Themes
Tags the research workers attached, counted across all findings (a finding can carry several tags; tags on fewer than 2 findings are left out). Not charted because nearly every finding has it: "mcp" (65 of 71).
Failures (41)
| Source date | Reported (our summary, first sentence) | Source |
|---|---|---|
| 2026-11-10 | In @ai-sdk/mcp over Streamable HTTP, a network-level failure (connection refused, connect timeout, DNS) on the 2026-07-28 server/discover probe is treated as a legacy server and falls back to initialize against the same unreachable… | github.com/vercel/ai#24817 |
| 2026-11-10 | In the Claude Code VS Code extension, remote HTTP MCP servers that accept unauthenticated connections but only work fully with a Bearer token cannot be signed in: the MCP panel shows Reconnect/Clear authentication/Disable/Remove but no… | github.com/anthropics/claude-code#108783 |
| 2026-11-10 | A user running the Todoist hosted MCP server (OAuth 2.1, rotating refresh tokens with reuse detection) across several parallel Claude Code sessions plus scheduled tasks reports the sign-in keeps getting revoked every few days. | github.com/anthropics/claude-code#108794 |
| 2026-11-09 | After the Codex app bundled a newer codex-cli, large stdio MCP replies (e.g. a base64 resources/read of a big binary file) exceed the rmcp transport's new default line-length limit and are silently dropped. | github.com/openai/codex#60383 |
| 2026-11-09 | In Codex CLI 0.167.0, when the first message of a session mentions a skill that declares an MCP server dependency, the 'Install MCP servers?' prompt is routed to the temporary title-generation thread, which rejects user interaction and… | github.com/openai/codex#60406 |
| 2026-11-09 | In Claude Desktop's Code tab on Windows, a third-party plugin's MCP App preview card fails when its confirm button issues a tools/call; the call is rejected client-side and never reaches the MCP server, while calling the same tool from… | github.com/anthropics/claude-code#108540 |
| 2026-11-09 | Remote MCP servers added as claude.ai connectors never refresh their tool list in a running Claude Code session after the server sends notifications/tools/list_changed (both 2025-era and 2026-07-28 delivery paths), while the same server… | github.com/anthropics/claude-code#108596 |
| 2026-11-09 | Authorizing the HeyGen MCP connector from Claude Code's /mcp fails every time with 'Invalid redirect URI'. | github.com/anthropics/claude-code#108607 |
| 2026-11-08 | CrewAI's MCP static tool filter compares configured names to sanitized tool names, so allow/block lists written with the names a server publishes (hyphens, camelCase) silently fail: an allow-list keeps no tools, and a block-list for a… | github.com/crewAIInc/crewAI#9248 |
| 2026-11-08 | Report that the reference mcp-server-git runs Git with repository configuration intact, so a pre-planted core.fsmonitor or post-checkout/pre-commit hook executes code when ordinary MCP tools such as git_status or git_checkout are called. | github.com/modelcontextprotocol/servers#5488 |
| 2026-11-07 | Mastra's MCPServer exposes agents as ask_<agent> tools, but the generated tool calls agent.generate without any memory thread or resource id, so a Memory-enabled agent served over MCP forgets everything between calls and MCP clients… | github.com/mastra-ai/mastra#29673 |
| 2026-11-07 | Codex 0.167.0 (CLI and desktop) auto-declines MCP form-mode elicitation requests without showing the user a form, so an MCP server that asks for write confirmation (here a CRM contact update) reports the user declined. | github.com/openai/codex#60158 |
| 2026-11-07 | Every 'claude mcp login' to a remote HTTP MCP server prints the SEP-2352 'no issuer stamp' warning even after a fresh successful login. | github.com/anthropics/claude-code#108158 |
| 2026-11-06 | A long local Codex conversation on Windows stayed in a working state for a long time with little progress: tool calls that each finished in seconds were separated by multi-minute idle gaps, and status questions got no response. | github.com/openai/codex#59849 |
| 2026-11-05 | In @mastra/mcp, tools rebuilt from cached definitions with toolsFromDefinitions() open their connection lazily on first execute(), but that handshake runs outside the call's request context, so a custom fetch that injects per-user… | github.com/mastra-ai/mastra#29531 |
Showing 15 of 41, newest source first. All of them: /api/trends?items=1.
Missing tools (10)
| Source date | Reported (our summary, first sentence) | Source |
|---|---|---|
| 2026-11-10 | Feature request: Claude assigns MCP server icons heuristically from tool names and offers no way for server authors or users to set a custom icon. | github.com/anthropics/claude-code#108788 |
| 2026-11-07 | Feature request for the OpenAI Agents SDK: a hook to compare an MCP server's listed tool definitions against an approved baseline (observe/review/enforce) because servers can change descriptions or schemas without a version bump, a… | github.com/openai/openai-agents-python#5588 |
| 2026-11-04 | Feature request for project-specific connectors: the user wants a domain-specific MCP connector available only inside one Claude project (and the same MCP authorized with different scopes per project) rather than account-wide in every chat. | github.com/anthropics/claude-code#107123 |
| 2026-11-04 | A user asks for connectors scoped to a single claude.ai Project. | github.com/anthropics/claude-code#107123 |
| 2026-11-01 | A Claude Code user (2.1.316, Linux) asks for support for skills delivered over MCP. | github.com/anthropics/claude-code#106550 |
| 2026-10-30 | An MCP tool call that continues as an MCP task can ask for input, for example an approval. | github.com/mastra-ai/mastra#28986 |
| 2026-10-29 | Plugins in Anthropic's own marketplaces bundle many MCP servers. | github.com/anthropics/claude-code#105932 |
| 2026-10-28 | @ai-sdk/mcp 2.0.86 rejects 2026-07-28 tool results with resultType 'input_required', even when an elicitation handler is registered. | github.com/vercel/ai#24153 |
| 2026-10-28 | A Codex app user asks Codex to discover and call WebMCP tools that webpages register in the user's connected Chrome tabs. | github.com/openai/codex#57853 |
| 2026-10-28 | Account-level claude.ai MCP connectors (Google Drive, Coupler.io, Supermetrics, Canva, semrush) are loaded into every Claude Code CLI and VS Code session, even when only needed on the web. | github.com/anthropics/claude-code#105769 |
Wishes for a directory (3)
| Source date | Reported (our summary, first sentence) | Source |
|---|---|---|
| 2026-11-06 | Users of the official MCP Registry UI report that search does not match server descriptions, so servers can only be found by name or exact keyword. | github.com/modelcontextprotocol/registry#1807 |
| 2026-11-05 | inferred: The census found that registry-declared auth headers did not match whether servers actually demanded credentials, and it names a tool picker that filters on the declared-headers field as a failure case. | digitalapplied.com/blog/mcp-remote-server-health-census-november-2026 |
| 2026-10-30 | A directory operator asked for mcp.market to be listed under Resources in the MCP servers repo. | github.com/modelcontextprotocol/servers#5415 |
Trends (12)
| Source date | Reported (our summary, first sentence) | Source |
|---|---|---|
| 2026-11-11 | The MCP reference servers repo shipped v2.4.0 with security fixes for server-filesystem (symlink/read races, move_file), mcp-server-git (repository Git config can no longer make the server run a program; tighter --repository… | github.com/modelcontextprotocol/servers/releases/tag/v2.4.0 |
| 2026-11-10 | IBM and Microsoft announced Imagine Enterprise, an agent platform for building, running and governing industry-specific AI agents (initially financial services, insurance and retail), combining IBM Consulting's prebuilt agents with… | news.microsoft.com/source/2026/11/10/ibm-and-microsoft-announce-a-new-ai-agent-platform-designed-for-industry-specific-transformation |
| 2026-11-10 | At Adobe MAX 2026 Adobe announced native MCP support so third-party AI agents can work on native project files in Photoshop, Illustrator, Premiere and After Effects (preview), with layers and sequences staying editable; the integrations… | creativeboom.com/news/adobe-max-2026-all-the-big-announcements |
| 2026-11-10 | browser-use released 1.0, rebuilding the agent so the model writes Python in a persistent REPL directly on Chrome DevTools Protocol. | github.com/browser-use/browser-use/releases/tag/1.0.0 |
| 2026-11-09 | Adversa's November MCP security digest summarizes research and advisories: a study scanning MCP servers across registries for undeclared filesystem capabilities produced coordinated advisories; it also covers cross-origin env… | adversa.ai/blog/top-mcp-security-resources-november-2026 |
| 2026-11-08 | The same source matched vendor domains against the ChatGPT and Claude connector directories and found coverage is fragmented, with many vendors listed in only one directory; matching was by domain only so the source calls the numbers a… | bigideasdb.com/mcp-use-cases |
| 2026-11-05 | In the same Nov 5 registry census, adoption of the 2026-07-28 spec's stateless server/discover method was low among sampled streamable-HTTP endpoints; most responding servers still negotiated the older initialize handshake. | digitalapplied.com/blog/mcp-remote-server-health-census-november-2026 |
| 2026-11-05 | A census of remote endpoints sampled from the official MCP Registry on Nov 5, 2026 sent each an unauthenticated initialize request: 550 completed the handshake, 338 asked for credentials, and 112 failed (DNS/TLS, timeouts, non-MCP… | digitalapplied.com/blog/mcp-remote-server-health-census-november-2026 |
| 2026-11-02 | The same MCP Bits #6 issue reports that the Ruby MCP SDK reached Tier 1, making six Tier 1 SDKs. | mcpbits.substack.com/p/mcp-bits-6 |
| 2026-11-02 | MCP Bits #6 (Nov 2, 2026) reports that SEP-3398 merged: the MCP spec now ships twice a year, on the last Tuesday of March and of September, with freezes and SDK betas before each. | mcpbits.substack.com/p/mcp-bits-6 |
Showing 10 of 12, newest source first. All of them: /api/trends?items=1.
Use cases (5)
| Source date | Reported (our summary, first sentence) | Source |
|---|---|---|
| 2026-11-09 | UC Today's weekly roundup describes enterprise agent deployments: Zendesk employee-service agents answering access, policy, onboarding and IT requests across Slack and Teams, and Box Architect (launched Nov 5) for building agentic… | uctoday.com/this-week-enterprise-ai-gets-serious-about-action |
| 2026-11-09 | CreatorDB describes marketing teams using its MCP server with Claude to build in-house influencer tools instead of buying SaaS: discovery agents that turn a campaign brief into a ranked creator shortlist, vetting tools, competitor… | manilatimes.net/2026/11/09/tmt-newswire/pr-newswire/creatordb-api-mcp-data-infrastructure-for-the-new-creator-economy/2465475 |
| 2026-11-09 | A perfume retailer (10ML) launched a keyless remote MCP server over its fragrance index so assistants like Claude, ChatGPT and Cursor can answer perfume questions; it is listed in the official MCP Registry and Claude's connectors… | globenewswire.com/news-release/2026/11/09/3398379/0/en/10ml-launches-a-free-perfume-api-and-perfume-mcp-server-built-on-its-fragrance-index.html |
| 2026-11-08 | A vendor lists 20 non-DevOps research jobs run through its MCP server in Claude and ChatGPT, with prompts and outputs: validating startup ideas against complaint data, competitor review teardowns, pricing SaaS acquisitions, ranking… | bigideasdb.com/mcp-use-cases |
| 2026-11-04 | In HN's 'Who wants to be hired? (November 2026)' thread, a developer lists PHI Guard MCP, a local-first MCP tool that scans for sensitive health information. | news.ycombinator.com/item |
Remote MCP servers: can an agent reach them? Measured
Every night we send each listed remote MCP server one unauthenticated server/discover or initialize request. We never send credentials and never call tools. Latest result for 809 servers:
Runs so far
| Day (UTC) | Reachable | Needs credentials | Failed |
|---|---|---|---|
| 2026-11-11 | 442 | 349 | 20 |
| 2026-11-12 | 0 | 1 | 0 |
2 run days so far: too few to call a trend. This table grows nightly.
By category
Categories with at least 5 checked servers. Categories are assigned by our public keyword rules, so they are inferred, not reported by the publisher. Green reachable, yellow needs credentials, red failed.
| Category | Checked | Reachable | Needs credentials | Failed | |
|---|---|---|---|---|---|
| AI & agents | 188 | 116 | 67 (36%) | 5 | |
| Other | 103 | 72 | 29 (28%) | 2 | |
| Search & research | 95 | 61 | 32 (34%) | 2 | |
| Payments & finance | 76 | 47 | 26 (34%) | 3 | |
| Productivity & documents | 59 | 27 | 32 (54%) | 0 | |
| Media & design | 53 | 19 | 34 (64%) | 0 | |
| Communication | 48 | 16 | 32 (67%) | 0 | |
| Analytics & marketing | 47 | 11 | 34 (72%) | 2 | |
| Commerce, travel & location | 45 | 29 | 15 (33%) | 1 | |
| Code & developer tools | 29 | 15 | 12 (41%) | 2 | |
| Cloud & DevOps | 25 | 10 | 15 (60%) | 0 | |
| Databases & data | 16 | 7 | 9 (56%) | 0 | |
| Security | 13 | 8 | 5 (38%) | 0 | |
| Browser automation | 12 | 4 | 8 (67%) | 0 |
Method and limits
- Reported by research workers: public sources found and summarized by AI research runs; counts of findings, not of users or traffic; not independently verified. Each finding has a public source link, a type and tags chosen by the research run. Duplicates (same URL and summary) are merged; 0 findings so far were found again in a later run. Test rows are excluded. Ingested 2026-11-04 to 2026-11-11.
- What findings are not: traffic, usage, market share or a survey. A theme with more findings means more public reports were found, which also depends on where the research runs looked (mostly GitHub issues).
- Measured: one unauthenticated server/discover or initialize probe per listed server per night. Labels: reachable, needs credentials (401/403, with the OAuth protected-resource metadata checked once), failed (DNS, TLS, timeout, HTTP error, not MCP). Full method: directory methodology. Not a guarantee a server works for you.
- Percentages are rounded shares of the counts shown next to them. Nothing on this page is estimated, modeled or extrapolated. Rankings and listings are never for sale.
Data: /api/trends (compact) · with every finding · MCP server uptime and tool trends.