Netcafe Memory MCP server

ai-netcafe

Hosted tools your agent can call over MCP — and a claim you can check in one command:

curl "https://ainetcafe.com/t/china_reachability?url=https://github.com"

{
  "url": "https://github.com",
  "tested_from": "mainland China (China Mobile backbone)",
  "verdict": "unreachable: timeout (common signature of blocking or a very slow route)"
}

That request was answered by a machine actually sitting on a Chinese backbone network. A monitor hosted anywhere else can only tell you GitHub is up where it is — which is not the question you were asking. That is the kind of thing we host: capabilities that are hard to get, rather than wrappers around ones that aren't.

Four focused servers

One agent profile rarely needs both spreadsheet cleanup and the Chinese lunar calendar, and every tool description costs tokens on every session. So the catalogue is served as four focused MCP endpoints — connect the one you need and pay the context cost of nothing else.

All four are in the official MCP registry as com.ainetcafe/netcafe-tables, -devkit, -docflow and -china.

Nothing in these four calls a language model or a third-party API. They compute on our own servers, so results are deterministic and repeatable — and they cannot break because someone else changed their API.

Numbers come with their own proof

Anything involving counts or money returns an arithmetic check computed in code, not asserted by a model:

// reconcile_ledger: your books vs the bank statement
{
  "total_a": 105.30, "total_b": 106.80, "gap": -1.50,
  "only_in_a":  [{ "key": "B",  "amount": 5.00 }],
  "only_in_b":  [{ "key": "C",  "amount": 7.00 }],
  "mismatched": [{ "key": "A3", "a": 100.00, "b": 99.50, "diff": 0.50 }],
  "arithmetic_check": {
    "formula": "A 105.3 − B 106.8 = -1.5; differences explain -1.5",
    "checks_out": true
  }
}

If checks_out is false, the response says the result is untrustworthy instead of handing back a table that looks finished. Amounts are compared in integer cents, so 0.1 + 0.2 never invents a difference for someone to go chase.

Connect it

claude mcp add --transport http netcafe-tables https://ainetcafe.com/mcp/table

DeepSeek Harness:

dsh plugin --profile web add github:mario03690/dsh-netcafe

Everything is also a plain GET — no MCP client, no SDK, no key:

curl "https://ainetcafe.com/t/cn_holiday?date=2026-10-01"
curl "https://ainetcafe.com/t/diff_text?a=one%0Atwo&b=one%0Athree"

Charts come back as images you can paste straight into a README:

That image is rendered on request, right now, by the same service this page describes.

Model cost is measured, not quoted

npx ai-netcafe compare "dedupe a JS array preserving order"

MODEL              COST       LATENCY
deepseek-v4-flash  $0.00012    7.6s
gemini-3.5-flash   $0.000341   10.0s
GLM5.2             $0.001828   9.0s

Shortened. The full README is on GitHub.