MCP · flash models · no Anthropic

The main model
should not read
the dump.

Ultra-Eye is an MCP server. You ask a question. A quota-plan flash model reads the files, the test log, the page, or the screenshot. You get a short answer with verified pointers or quotes — and a receipt. The raw dump never enters the main model’s context.

[Ultra-Eye · read 11,497 tokens → returned 181 · flash 10,520 in / 181 out · glm-5.3-flash · 7.9 s]

The problem

A multi-file question or a 50k-line pytest log should not be ingested by the model you pay for. That is quota spent on reading, not on deciding.

What Ultra-Eye does

It reads on OpenCode GO and GLM Coding Plan flash models. You keep the answer, the quotes, and the receipt. Judge a call from read versus returned.

How a call works

You ask

One question. A folder, a command, a URL, or an image. The main model does not open the files.

The eye reads

Flash models on the quota plan search, read, fetch, or look. Tiny scopes skip the model and return raw.

You get a slip

A short answer, disk-true quotes or pointers, confidence, and one receipt line. Repeat that line at the end of your reply.

Five tools

eye_task

Delegate a multi-file question or a long command. Inner loop of search / read / list / run / fetch. One to three minutes. Do not poll.

eye_scan

Single-shot codebase lookup. Direct answer plus verified file:line pointers quoted from disk.

eye_run

Run a command and triage large output. Exit, verdict, verbatim failures, path to the raw log. Errors are copied, never paraphrased.

eye_fetch

GET a page and answer a question. Quotes are the model’s selections, validated against the extracted text. No padding, no login.

eye_look

Mechanical read of a local PNG, JPEG, or WebP: text in a screenshot, presence, error-dialog copy. Not layout or design judgement.

The receipt

Last line of every response. Same numbers as the log and python -m ultraeye.report. The word “saved” does not appear — avoided reading is not saved quota.

read
Raw material the eye processed: file text, extracted page, command output. For eye_task, the sum of inner observations.
returned
Token estimate of the caller-facing body, including quotes.
flash in / out
The eye’s own model usage on the quota plan. No model: flash 0 in / 0 out · no model.
steps
Appended on eye_task only: · 15 steps.

Install

1. Clone and venv

git clone https://github.com/Zhihong0321/Ultra-Eye.git
cd Ultra-Eye
python -m venv .venv
.venv\Scripts\python.exe -m pip install -e .
copy ultraeye.example.json ultraeye.json

Set vault_path in ultraeye.json to your credential store. Credentials are read by name, never shipped.

2. Point an agent at stdio

python ultraeye/server.py

Windows-native. Version 1.0.0. Then register the same command in the agent:

grok mcp add ultra-eye -- ".venv\Scripts\python.exe" "ultraeye\server.py"
grok mcp enable ultra-eye

3. Tell the agent

Paste AGENT-GUIDANCE.md into the project AGENTS.md or CLAUDE.md.

  • Several files, or more than a screen of output → one eye_task.
  • A page or an image → eye_fetch or eye_look.
  • A lookup one grep would answer → do it yourself.
  • Expect one to three minutes. Do not poll.
  • Repeat the receipt line at the end of your reply.

What it will not do