WRONG DATA, SILENTLY
Postgres MCP Pro (crystaldba)
Columns that share a name are silently dropped
A query selecting count(distinct city), count(*) and count(city) came back as [{'count': 69}]: two of the three values were gone with no warning. Results are keyed by column name, so duplicate names overwrite each other. Rows are also returned as Python repr, not JSON.
Agents: Claude Code hit it in 3 of 5 attempts; Codex aliased its columns and never did.
Fix: Return rows as arrays with a separate column list, or de-duplicate column keys (count, count_2…).
WRONG DATA, SILENTLY
DBHub (Bytebase)
Columns that share a name are silently dropped
count(distinct city), count(*) returned a single {"count": "61"}: the agent reported the total row count as the number of distinct cities. Same root cause: result rows are objects keyed by column name.
Agents: Claude Code hit it in both attempts; Codex aliased its columns.
Fix: Preserve duplicate columns (array rows, or suffixed keys).
WON'T START
Postgres MCP Pro (crystaldba)
Latest release crashes on startup with the current MCP SDK
postgres-mcp 0.3.0 imports mcp.server.fastmcp, which mcp 2.x removed. A fresh install pulls mcp 2.x and the server exits immediately. It works when mcp<2 is pinned.
Agents: All agents (the server never starts).
Fix: Pin mcp<2 in the package metadata, or migrate to the 2.x API.
WON'T START
MCP Toolbox for Databases (Google)
No Linux arm64 build (Docker on Apple silicon, AWS Graviton)
The npm wrapper only ships darwin and linux-x64 binaries and fails with "Unsupported platform: linux-arm64". We ran Google's x64 binary under emulation; everything passed once it ran.
Agents: All agents on ARM Linux.
Fix: Publish a linux-arm64 binary and npm optional dependency.
WON'T START
Lightpanda
Binary requires glibc 2.38: fails on Debian 12 and Ubuntu 22.04
The 1.0.0 aarch64 release exits with "GLIBC_2.38 not found" on the two most common server distros. It ran fine in a Debian 13 container.
Agents: All agents on older Linux distros.
Fix: Build against an older glibc, or ship a static/musl build.
HANGS
Fetch (official)
Every HTML fetch hangs in sandboxed or offline environments
The HTML-to-Markdown step (readabilipy) installs npm packages the first time it runs. Without internet access, as in most agent sandboxes, every HTML fetch hung until the 180-second tool timeout. JSON and plain-text fetches were fine.
Agents: All agents, in any network-restricted sandbox.
Fix: Vendor readabilipy's JS dependencies at install time, or fall back to the Python parser when Node deps are missing.
WRONG DATA, SILENTLY
open-webSearch (fetchWebContent)
includeLinks: true returns link text mashed together, with no URLs
A page with five links came back as "denamabhuhebhoresusdobizubcatub": the five link labels concatenated, no spaces, no hrefs. Agents can't follow links or report where they point.
Agents: Claude Code and Codex both failed every link task (0 of 8).
Fix: Emit links as markdown [text](href) when includeLinks is set.
MISSING CAPABILITY
DuckDuckGo MCP (fetch_content)
Fetched pages lose every link URL
fetch_content in markdown mode returns link labels as a bullet list with the URLs stripped. The alternative backend needs an optional extra that the default install doesn't include ("pip install 'duckduckgo-mcp-server[browser]'").
Agents: Claude Code and Codex both failed every link task (0 of 8).
Fix: Keep hrefs in markdown output; make the error message part of the docs.
BREAKS WITH SOME AGENTS
fetcher-mcp
Long pages work with Codex, fail with Claude Code
fetch_url returns the whole page in one response and offers maxLength but no offset. Claude Code caps tool output (~25k tokens), so content past the first ~20,000 characters is unreachable. Codex accepted the full 380k-character response and passed.
Agents: Claude Code 0 of 2; Codex 2 of 2.
Fix: Add an offset/start_index parameter (as the official fetch server does).
FAILS ON SLOW SITES
mcp-searxng (web_url_read)
Hard-coded 10-second page timeout
web_url_read gives up after 10 s ("took longer than 10000ms to respond"). Retrying never helps, so any slow page is unreadable.
Agents: Claude Code and Codex both 0 of 2.
Fix: Make the timeout configurable and default it higher (e.g. 30 s).
BREAKS WITH SOME AGENTS
Playwright MCP (ExecuteAutomation)
Screenshots are saved to disk, not shown to the agent
playwright_screenshot returns "Screenshot saved to: …/Downloads/page-….png" instead of the image, so Claude Code could not see what was on the page.
Agents: Claude Code 0 of 2; Codex 2 of 2.
Fix: Return the image content in the tool result by default.
MISLEADING OUTPUT
browser-use (MCP mode)
A missing page comes back as an empty page
Navigating to a 404 returned "No content extracted" with the URL as the title: no status code or error text, so agents can't tell "missing" from "empty". (Claude then claimed "404" without evidence; our verifier rejects claims the tools didn't show.)
Agents: Claude Code 0 of 2.
Fix: Surface the HTTP status and error page text.
SETUP FRICTION
@agent-infra/mcp-server-browser (ByteDance)
"Unable to find any browser" on Linux arm64
It only looks for a system Chrome, which has no Linux arm64 build. It works once --executable-path points at Chromium; after that it passed every task with Claude Code.
Agents: All agents on ARM Linux, until configured.
Fix: Fall back to Playwright's bundled Chromium, or say how to fix it in the error.
BY DESIGN
Lightpanda
Doesn't render images; multi-step forms were flaky
Lightpanda is a fast non-graphical browser, so screenshots can't show image content. A three-step form wizard also failed 2 of 5 attempts with Claude Code.
Agents: Claude Code (wizard 3/5, screenshot 0/2).
Fix: Document the no-render tradeoff in the tool descriptions so agents pick another tool.
Completion rates per agent, with confidence intervals and statistical tie bands.
