cheap-eyes

Cheap eyes for an expensive agent.

An AI agent's context is expensive. When it reads a big log, a long doc, a web page or a whole code tree, it burns that context on lines it mostly doesn't need. cheap-eyes is an MCP server that hands the reading to a cheap model. It then checks every line of that model's answer against the source, so the agent gets back a short answer it can trust.

What it looks like

A real header from the server (trimmed), with a stubbed model answer. The agent asked for every ERROR line after 08:10 in an app log, plus the traceback that follows it:

eyes_run result: 2026-09-26_152246-extract-889090
model: alias fast -> qwen/qwen3.6-35b-a3b (used: qwen/qwen3.6-35b-a3b)
tokens in 412 / out 96; cost $0.000080; 2 ms
check (extract): refs ok 4 / bad 1 — ok 4, ok~ 0, partial 0, not-in-input 0
  L9 [bad] text differs from the source line
--- untrusted model output ---
L3| 2026-09-20 08:14:09.031 ERROR [sync] upload failed: timeout after 30 s
L4| 2026-09-20 08:14:09.032 Traceback (most recent call last):
L6| TimeoutError: timed out
L8| 2026-09-20 08:15:31.550 ERROR [sync] upload failed: HTTP 503 from api.example.com
L9| 2026-09-20 08:16:00.000 ERROR [sync] retry 2/3 failed
--- end ---

The model quoted five lines. In the real log, line 9 says INFO [sync] retry 2/3 ok, so the model made up an error. The check compared each quote with the source line and flagged L9 as bad. The agent sees that before it relies on the answer.

Related MCP server: delegate-mcp

Why you can trust the answer

• The model has no tools. It can't open files, browse or call anything. It sees only the numbered lines cheap-eyes chose to send.

• Secrets are masked before anything leaves your machine. Passwords, tokens and keys become [token], [key] and so on, in the files and in the task.

• Only Zero Data Retention providers. Those are providers that don't store your prompts. cheap-eyes talks only to OpenRouter and asks for them on every call.

• Every quote and line reference is checked mechanically against the lines that were sent. Wrong values and uncited claims are flagged.

• What the check does not prove: that the answer is complete or relevant. It proves what was quoted is real, not that nothing was missed.

Requirements

Installation

Claude Code / Claude Desktop (npx)

1. Write a config file at ~/.config/cheap-eyes/config.json, or %APPDATA%\cheap-eyes\config.json on Windows (see Configuration).

2. Check it (no network, no secrets printed): npx -y cheap-eyes check-config

3. Add the server to Claude Code:

   claude mcp add --env CHEAP_EYES_OPENROUTER_KEY=<OPENROUTER_KEY> --transport stdio cheap-eyes -- npx -y cheap-eyes

4. Try it: "Use eyes_run to list every TODO in /home/me/notes/**/*.md, with the line it's on."

Claude Desktop, .mcp.json and a global install: installation.

Home Assistant app

Home Assistant renamed add-ons to apps in 2026.2; on older versions the same menus say Add-ons and Add-on Store.

1. Add this repository:

   

   Or by hand: Settings → Apps → App Store → ⋮ → Repositories → + Add, paste https://github.com/st412m/cheap-eyes, select Add.

2. Find cheap-eyes in the store and select Install.

3. In Configuration, set openrouter_key, an http_token of at least 32 characters (openssl rand -hex 32 gives one) and your read_roots.

4. Select Start. The app serves MCP on port 3400.

Options and the connector URL: app documentation.

Docker / HTTP

cheap-eyes serve --http runs the same server over Streamable HTTP with a token; the repository's Dockerfile and compose.example.yaml wrap it. See installation.

Configuration

A JSON config file. The fields to set first:

{
  "read_roots": ["/home/me/notes"],
  "models": { "fast": { "ids": ["qwen/qwen3.6-35b-a3b"], "params": { "reasoning": { "effort": "none" } } } },
  "defaults": { "extract": "fast", "draft": "fast", "edits": "fast" }
}

The key goes in the env var CHEAP_EYES_OPENROUTER_KEY, never in the file. Every field, env var and file location: configuration.

Connecting to claude.ai

Run the HTTP server (or the Home Assistant app) and publish it through a reverse proxy that terminates TLS. The endpoint is:

https://<your-host>/private_<your-token>/mcp

Add it in claude.ai → Settings → Connectors → Add custom connector. A wrong path gets a bare 404, and there is no OAuth by design. See security.

Using it

You talk to your agent as usual. It calls cheap-eyes when a job fits. Pick the mode by what you want back:

• extract when you need exact lines: errors in a log, a config value, every mention of something. Invented lines are flagged.

• draft when you want notes or a summary. Every line must cite its source lines. Values that aren't in those lines, and lines with no citation, are flagged.

• edits when you want fixes proposed. You get JSON edits whose old text is checked against the file. cheap-eyes never applies them.

• schema when you want fields filled: a JSON schema in, the same JSON out with a line ref and a verbatim quote for every value. Conflicting values are reported, not resolved.

• grep when you know the exact pattern: matching windows, no model, $0.

Files can be text, HTML, PDF (with a text layer), Word (DOCX, DOC), RTF, PowerPoint (PPTX, PPT), OpenDocument text and presentations (ODT, ODP), EPUB, FB2 and mail (EML, MSG, attachments included), or an http(s) URL that the server fetches itself. Spreadsheets and scans are refused with the reason. Refs name the page, slide, chapter or attachment of a line: L40 (slide 3), L12 (attachment: offer.pdf, p.2) (formats).

Prompts you might type (the files must lie inside your read_roots):

• "Use eyes_run to pull every ERROR from /var/log/app/app.log since 08:00 today, with the tracebacks."

• "Have cheap-eyes draft a summary of /home/me/notes/deploy.md: versions, ports, backup times."

• "Use eyes_run in edits mode to fix the dead links in /home/me/notes/*.md."

• "Use eyes_run in schema mode on /home/me/docs/tender.pdf: customer, maximum price, deadline, and the list of items with quantities."

What each check catches and misses: checks.

When to use cheap-eyes and when grep

cheap-eyes pays off for reading by meaning over large text: logs, archives, long documents, web pages, "everything about X", chronologies. For exact lookups — a key, which blocks contain X, a list by a field, a value in a table — and for files under about 30 KB, grep, a ranged read or code is cheaper and more reliable. That is why spreadsheets are not read at all.

On a benchmark of 2026-09-30 with independent ground truth:

• An 887 KB log archive, task "everything about power": recall 96 %, precision 100 %, $0.08, about 12k tokens of the agent's context instead of about 227k for reading it whole.

• A structural task over Home Assistant automations ("who sends to Telegram"): precision 41 % with a green check. The check catches invented lines, not wrongly chosen ones.

A snippet for CLAUDE.md or your Claude settings:

Reading by meaning over large text (logs, archives, long documents, web pages, "everything about X") → cheap-eyes eyes_run (extract/draft; fields → schema). Exact values, structure, small files (< ~30 KB) → grep / ranged read (no model: eyes_run mode grep). Verify any value from cheap-eyes against the source (eyes_result source) before relying on it.

To compare models on your own documents, run cheap-eyes bench.

Tools

• eyes_run — does the job and returns a short, checked header.

• eyes_result — reads a stored result: pages, grep, the check report, the source text the model was given; cancels a running job.

• eyes_stats — spend today and all time, your key's status; on request, the model table.

Inputs and outputs: tools.

Models and cost

No model is pinned in code. You name models in your config under aliases (fast, long, …), and switching means editing the config and restarting. Each alias lists model ids in priority order; if one can't be used, the next is tried. The example ids were valid on 2026-09-26 and will age. To pick new ones:

cheap-eyes models --candidates

This lists cheap Zero Data Retention models, and shows how well each of your aliases passed its checks on your own jobs. If your OpenRouter account's Guardrails don't allow a model, cheap-eyes skips it and marks it blocked by account. Allow it at openrouter.ai → Guardrails if you want it used.

With cheap models a job costs fractions of a cent. Your key's own daily limit on OpenRouter is the real ceiling. On top of that, daily_budget_usd (default $1 a day) stops calls before they would go over. More: models and spending.

Limitations

• The check proves quotes are real, not that the answer is complete. A relevant line the model skipped is not flagged.

• No OCR. Scanned PDFs are refused with the reason; layout and images are lost in extraction.

• No spreadsheets. Excel and OpenDocument spreadsheets are refused: questions about a table are exact lookups, which grep or code answers and a model reading the table as text gets wrong.

• Size caps. A text file (or the text extracted from a document) over 2 MB is refused, never truncated; a document over 50 MB and a URL over 10 MB likewise; a job is capped at 5 MB after filtering and 200 files.

• No reduce pass for draft. With several chunks, each chunk's notes come back under their own header.

• The hook covers Claude Code's Read tool only. Shell commands are not intercepted.

Troubleshooting

• OpenRouter answers 403 "Access denied by security policy". OpenRouter refuses some regions; route the calls through a proxy.

• A model is blocked by account. Allow it at openrouter.ai → Guardrails, or put another id in the alias.

• The server won't start and names an alias. A value in defaults is not defined in models.

• The client shows old tools after an update. Refresh the connector's tool list and start a new chat.

More: troubleshooting.

Security

• The model gets no tools; everything it returns is framed as untrusted output.

• Only files inside read_roots are read, after resolving symlinks; .env*, keys, secrets.yaml, databases and .git are always refused.

• Secrets are masked before anything leaves the machine; model calls go only to openrouter.ai, with Zero Data Retention required.

• URLs in a job are fetched by the server with private and local addresses refused, also against DNS rebinding.

• The server writes nowhere outside its own state folder unless you opt into export.

• In HTTP mode the token works as a password. It needs at least 32 characters and is never logged; put TLS in front.

More: security.

Links

• Documentation — installation, configuration, tools, checks, models and spending, results, security, hook, formats, bench, troubleshooting, internals

• Home Assistant app

• Changelog

Prior art

Both MIT-licensed; cheap-eyes is written from scratch (see NOTICE).

• BardyLaw/claude-worker-delegation — the server reads files by path, short preview, usage log, daily budget, PreToolUse hook. cheap-eyes adds what it lacked: a realpath path guard, refusal instead of silent truncation, masking, line numbers and checking the output.

• aashahin/claude-opencode-delegate — async task registry, compact status header.

License

MIT — by st412m.