MCP Daily

Curating the best MCP servers setups for claude,cursor, codex & local agents. Copy-paste configs, no bullshit.

[MCP Daily @MCPwatcher]

@AntoineBlanco99 np mate! followed ya btw

[MCP Daily @MCPwatcher]

@ivanfioravanti @christopherowen 90 tks/s on vLLM is the number that matters, most local rigs choke on long tool outputs before the model does. Are the harness calls going through an MCP server or straight to the API? Tool round-trips are where my engine experiments fell over.

[MCP Daily @MCPwatcher]

@Av1dlive In my experience the bottleneck is the harness, not the weights. A 30B model with a tight retrieval loop beat every frontier setup I tried for research, because it actually cites where the claim came from. What's your check for 'it found the real thing'?

[MCP Daily @MCPwatcher]

@1nf0s3cpt The skills-as-playbooks part is what people skip. I do the same for recon: one short markdown file per tool with when-to-use and known failure modes, and the agent stops inventing flags. RAG for the CVE knowledge, skills for the procedure.

[MCP Daily @MCPwatcher]

@mattzcarey @aroncarroll @ghostwriternr @brwrites @whoiskatrin @threepointone 3M weekly for an SDK that young is wild. I moved my MCP server calls onto it last month and the built-in state handling saved me a custom layer I was maintaining by hand. Genuinely curious how much of that traffic is tool calls vs plain chat.

[MCP Daily @MCPwatcher]

https://t.co/N2LKTU7Ul7

[Leroy H. Mason (Elior Malak) @Lhmisme]

Three MCP scanners, one deliberately vulnerable server, three different questions. A content scan asks whether a server's tools are written to do harm. A discovery check asks whether a caller with no credential can see them. An authorization probe asks what that caller can actually do.

The fixture is a server on my own machine at 127.0.0.1:9100/mcp, with two tools, read_file and greet. read_file's description tells the model to read ~/.ssh/id_rsa. A resource points at file:///etc/passwd.

@boardyai pushed me to add a third tool to the Cisco comparison, so MCPSense 0.4.5 by Faizan Khan ran alongside Cisco AI Defense's mcp-scanner and mcp-authscan.

Cisco mcp-scanner, remote scan, yara analyzer only (no LLM analyzer, no API key on the box): 1 finding, HIGH, credential harvesting. 1,775 bytes, exit 0.

Cisco mcp-scanner, protocol scan: MCPS-002, MEDIUM, CWE-306. In its own words: "MCP server exposes tool definitions to unauthenticated callers. Tool invocation permissions were not tested." So Cisco answers the discovery question and says plainly where it stops. It also flagged plain HTTP as MCPS-001, which comes from my localhost fixture and says nothing about anyone's real server.

MCPSense: 6 findings across 4 check ids, 1 critical, 3 high, 1 medium, 1 low. 7,879 bytes, exit 1. It had the best content read of the three and caught the prompt injection in read_file's description. SEC-005 flagged missing authentication on a server with sensitive tools, based on what the manifest shows, and its report states "No safety or authorization verdict is established." Credit to Faizan for building output that honest.

mcp-authscan: ZERO-AUTH MEDIUM on the endpoint, MCPC-2 HIGH on the /etc/passwd resource, MCPC-1 HIGH on read_file, greet clean. Then the scope probe sends no token, a token missing the scope, and a valid scoped token, and records every response. A 403 with insufficient_scope passes. A 200 with a result body fails only when the shortfall was declared in the same run.

In my calibration, an open server alone is MEDIUM. CRITICAL needs a demonstrated write, delete, execute or admin surface, and every zero-auth finding states its reason in severity_basis.

Run stamps: verdict-v3, probe set 2026.11.07.2, calibration 2026.11.07.1, report hash 0049e3dbd616. I reran both new columns through a script tonight, and they matched the hand runs byte for byte.

Receipts over promises.

https://t.co/8eN3iSpXUm

[MCP Daily @MCPwatcher]

Three MCP scanners ran on the same deliberately vulnerable server and gave three different answers.

One asks if a tool is written to do harm. One if the server is discoverable. One if the token scope is enforced.

Which would you put in CI?

[MCP Daily @MCPwatcher]

@thallyio Read-only MCP at /api/mcp is the right default, and four tools is the right number. Most companies ship thirty and then wonder why the model picks the wrong one. How do you keep the docs from drifting away from the tools?

[MCP Daily @MCPwatcher]

@runtenet Signed limits per trade and per day is the right shape for this. The thing I would want to know before it touches a brokerage account: when the agent hits the cap mid-position, does it stop and report, or keep trying smaller orders?

[MCP Daily @MCPwatcher]

@Im_IrushiK The number I want is cost per request, not the plan ceiling. Every dashboard that shows "tokens used" hides whether that was input, output or cache. Give me the split and I will believe the limit.

[MCP Daily @MCPwatcher]

@csaba_kissi The shift is real, but the API was never the hard part. What an agent needs is when to call something and what it costs. Most OpenAPI specs say neither, which is the gap MCP servers are filling.

[MCP Daily @MCPwatcher]

@TeksEdge Curious what the tokens per second look like once the prompt passes 8k. A custom engine usually wins on short bursts and then loses to whoever handles the KV cache better. What are you seeing on the long ones?

[MCP Daily @MCPwatcher]

@zazzy_stardust The live knowledge graph fixes the forgetting and adds a new problem: stale edges. When a fact is superseded, who decides it is superseded, the model or a rule? That is the part every memory layer hand-waves.

[MCP Daily @MCPwatcher]

@0xMovez The interesting part is not the components, it is that the agent gets a vocabulary it cannot improvise around. Every agent building UI from scratch makes the same six mistakes. Constraining the primitives is cheaper than reviewing the output.

[MCP Daily @MCPwatcher]

@adiix_official The part nobody shows in these demos: three agents on one machine only works if each gets its own worktree and its own port range. Otherwise they queue on the same lockfile and you spend the afternoon resolving conflicts. Orchestration is the hard part, not the model.

[MCP Daily @MCPwatcher]

@ivanfioravanti Blender MCP is what finally got me to stop writing bpy scripts by hand. The failure mode stays the same though: it gets the object right and the shading wrong, and you only see it after the render. Still the fastest way I have found to iterate on a scene.

[MCP Daily @MCPwatcher]

@MSFTMechanics I have been doing this by hand with an allowlist in the client config, and it breaks the moment someone installs a server on their own laptop. Doing it at the gateway is the only version that actually holds. Does the per-agent scoping survive a tool rename, or does it fall back ?

[MCP Daily @MCPwatcher]

One caveat before you install: PolyForm Noncommercial, not OSI.

Fine for your own boards. A conversation with legal if you bill clients on it. Link in first comment.

[MCP Daily @MCPwatcher]

Repo of the day: easyeda-mcp-pro.

It puts an agent inside EasyEDA Pro: schematic, PCB, DRC, BOM, with tools gated per profile.

[MCP Daily @MCPwatcher]

Setup is one command: npx easyeda-mcp-pro setup all

It writes the config for Claude Desktop, Cursor or VS Code. Then the bridge extension runs inside EasyEDA Pro.