unrent

by stringcutter

Find the strings your AI code hangs by. Cut them before they snap.

• Closed AI services the code calls, with file and line, and the open source that replaces each one, ranked.
• Models that stop working: model ids the code selects that their vendor has retired, or will on an announced date, and the vendor's replacement.
• Open source AI the code already runs, and where it ranks in its field.

The CLI runs offline. No account. No upload. No telemetry.

Real output. One row per string, with its strongest location and how many more there are:

--why shows every line behind one row, and what replaces it:

In a pipe or a file the same scan is a Markdown report that pastes straight into an issue or PR.

Install

uv tool install unrent      # or: pipx install unrent
unrent .

Or once, without installing:

uvx unrent .

Python 3.11+. One dependency: PyYAML (the MCP server adds mcp). Large repos are read on every core. With ripgrep on PATH it is about twice as fast again: 6 s for 5,500 files, 14 s for 14,000 (13 s and 26 s without, on 4 cores). Same results either way.

Use

unrent .                               # the terminal view; Markdown in a pipe or file
unrent . --why openai                  # every location of one service, and what replaces it
unrent . --why gpt-4-turbo             # every line that selects a retiring model
unrent . --as-of 2026-12-01            # judge retirements as of another day
unrent . -o report.md                  # the Markdown report
unrent . --format json -o r.json       # everything, machine-readable
unrent . --top 5                       # more alternatives per kind
unrent . --skip-tests                  # ignore test, spec and fixture code
unrent . --exclude "examples/"         # .gitignore syntax; or a .unrentignore file
unrent catalog                         # what it knows

unrent . is short for unrent scan .. Colour follows NO_COLOR and FORCE_COLOR.

On Windows PowerShell, write reports with -o, not >. The redirect re-encodes the file.

MCP

unrent as tools for Claude Code, Cursor, Copilot, or any agent that speaks MCP. The agent gets the facts; you decide what to swap.

claude mcp add unrent -- uvx --from "unrent[mcp]" unrent mcp

Other clients:

{
  "mcpServers": {
    "unrent": {
      "command": "uvx",
      "args": ["--from", "unrent[mcp]", "unrent", "mcp"]
    }
  }
}

Rankings come from this repo's main, refreshed weekly, not from the install. Cached for six hours. When the fetch fails it uses the shipped snapshot and says so. UNRENT_OFFLINE=1 never fetches. Only rankings come in. Your code never goes out.

For agents that use skills, npx skills add stringcutter/unrent adds a workflow on top: check what unrent can't see, sweep for services outside the catalog, and write the report.

What it detects

357 closed AI services in 21 categories: LLM APIs, LLM gateways, embeddings, vector databases, RAG platforms, document parsing, guardrails, classification and scoring, LLM observability, model hosting, fine-tuning, speech-to-text, text-to-speech, voice agents, image generation, web search, web scraping, browser automation, code sandboxes, agent platforms, agent memory. 218 are written by hand; 139 hosted model providers come from models.dev and are regenerated every week (models-dev.yaml).

Models that stop working. 204 model retirements announced by OpenAI, Anthropic and Google for their own APIs, with the date and the vendor's replacement, checked every week against their deprecation pages (retirements.yaml) and shipped with each release. A model id counts when a line selects it: a default, a config value, a model passed to a call. The same id in a model menu, a price table or a check on what the user picked is only counted. Azure OpenAI, Bedrock and Vertex keep their own schedules and are not covered. On the golden corpus: precision 0.939, recall 0.886.

Services it does not know yet. Every scan also lists unknown_candidates: API hosts and keys that no catalog entry explains and that look like a hosted AI API (a /v1/... path, an api. or .ai host, a matching *_API_KEY). They are candidates to check, not findings, and never counted.

89 open source projects: vector databases, inference servers, gateways, RAG frameworks and applications, document parsers, observability, evals, speech, scraping, agent memory.

Evidence it reads:

Every package name in the catalog exists in its registry. Checked weekly.

What it ignores

• Local servers behind a compatible SDK. OpenAI(base_url="http://localhost:11434/v1") is Ollama, not OpenAI.
• Lockfiles. Your dependencies' dependencies aren't yours.
• Comments and docs. Per-language comment syntax. "/api/*" in a string is not a comment.
• Ignored and vendored files. .gitignore with or without git. Submodules scanned. node_modules not.
• Generic names. task="transcribe" isn't Amazon. import textract isn't AWS. Your perplexity.py isn't Perplexity.
• Open-weight models. gpt-oss, deepseek-v3.2 and Ollama tags aren't closed.

Also:

• Secrets are masked. OPENAI_API_KEY=****.
• Test-only findings are marked. --skip-tests drops them.
• Model names alone don't count. A model id with no SDK, key, host or package behind it is listed apart, as closed models named in code.
• Vendor through vendor. Azure via openai, Claude on Bedrock or Vertex via anthropic: the report names the one you actually call.
• Oversized files are listed, never skipped silently.

Every rule above has a test that fails without it.

Ranking

Pools live in catalog/alternatives.yaml. Re-ranked every Monday by a GitHub Action that merges itself.

• Projects: GitHub stars gained in the last 90 days, from unrent's own weekly snapshots. Until four weeks exist: total stars. The report says which.
• Models: Hugging Face trending. Labs' own releases only, open licence, max two per lab.
• Open source only. OSI licence for code. Apache, MIT, BSD or CC-BY for weights. Restricted weights are out, whatever the code licence. Archived projects and a year without a push are out. Open core is marked: LiteLLM's enterprise/ or Langfuse's ee/ is not under the licence shown.
• No silent rot. A renamed, archived or relicensed project blocks the auto-merge and waits for a human. A pool that empties or halves is not published.

The scanner reads the snapshot shipped with the release. New rankings come with new releases.

Contribute

The catalog is YAML. New service: catalog/services/. New alternative: a pool in catalog/alternatives.yaml. False positive or miss: open an issue with the line that fooled it. CONTRIBUTING.md.

Licence

Apache-2.0. unrent is made by stringcutter.