Rango

Chameleon Ultra AI Copilot — read, crack, dump, analyze, emulate and clone RFID cards through natural language, with safety gates and a host-side card library.

Prerequisites · Quick start · Other harnesses · MCP tools · Coverage · Development

Chameleon Ultra — image © RfidResearchGroup

Rango turns a Chameleon Ultra into an agent-driven RFID copilot. It is only the AI layer: the official Chameleon Ultra CLI is not vendored into this repo — clone it separately; Rango locates and drives it in-process (see Prerequisites).

Two ways to use it:

1. Kimi Code plugin (best experience) — an MCP server + skill + analysis subagent; the Kimi Code agent operates the device for you.

2. Standalone AI shell — an enhanced REPL around the stock CLI where ? talks to an LLM (Moonshot/Kimi API or any OpenAI-compatible endpoint).

Architecture

flowchart LR
    U(["You — natural language"]) --> A["Agent harness<br/>Kimi Code · Claude Code · Codex · OpenCode"]
    SK["SKILL.md<br/>workflows · safety gates · coaching"] -. guides .-> A
    A -->|"MCP over stdio"| M["chameleon MCP server<br/>16 tools · one persistent serial owner"]
    M --> S["ai_shell<br/>bridge · ops · library · dictionaries · analyze"]
    S --> C["official Chameleon Ultra CLI<br/>fetched · unmodified · not vendored"]
    C -->|"USB serial"| D(["Chameleon Ultra"])

Related MCP server: MCP-Edge

Prerequisites

git clone https://github.com/fedroraddict/rango rango && cd rango

scripts/install-cli.sh   # fetches the stock upstream CLI into ../ChameleonUltra
                         # (git clone, with a codeload tarball fallback when
                         #  github.com is unreachable)
uv sync
scripts/setup-plugin.sh  # writes plugin/.rango-root so the MCP launcher still
                         # finds this repo after /plugins install copies plugin/

The installer is idempotent and prints what it did. If you keep the upstream checkout somewhere else, pass the path (scripts/install-cli.sh /path/to/ChameleonUltra) and set export CHAMELEON_SOFTWARE=/path/to/ChameleonUltra/software.

Requires Python ≥ 3.10, uv, and the device on a data-capable USB-C cable.

Quick start

# Standalone AI shell (stock CLI commands pass through; '? ...' asks the AI)
cd rango && uv run python -m ai_shell
#   needs MOONSHOT_API_KEY (or another OpenAI-compatible endpoint in
#   ~/.chameleon_ai/config.toml) for AI features; plain CLI works without it

# Kimi Code plugin
#   in Kimi Code: /plugins install <this-repo>/plugin   then  /reload

Use from Codex, Claude Code, OpenCode

The MCP server is a standard stdio server — not Kimi-only. Per-harness config snippets and install notes live in adapters/: Claude Code (.mcp.json / claude mcp add + drop-in skill and agent), Codex CLI ([mcp_servers.chameleon] + skill with stripped frontmatter), OpenCode (opencode.json + converted agent file).

MCP tools (16)

What the copilot layer adds

• Pre-flight ritual — workflows start with chameleon_state; gated steps are announced first.

• Detection coaching — failed scans trigger placement guidance, HF/LF alternation, app cross-check — not silent retry loops.

• Key-recovery decision tree — default dictionary → targeted dictionary built from a web search of the card system's known keys → hf mf autopwn → manual PRNG attacks (darkside / nested / senested / hardnested, with expected durations) → last resort: mfkey32v2 reader-side recovery (hf mf elog --decrypt).

• Library-first model (mirrors the CU GUI's Saved Cards): dumps get user-chosen names in the library; slots are working memory.

• Slot discipline — list first, prefer free slots, ask before overwriting, always name, always enable ((disabled) slots don't emulate), persist with hw slot store.

• Collaboration modes — copilot (confirm each gated step) or autopilot (approve a stated workflow once). hw dfu, hw factory_reset, and physical-card writes always get their own confirmation.

• Offline analysis subagent (card-analyst) for deep dump inspection.

Card coverage

Known quirks (verified on hardware)

• hf mf fchk --dic is a no-op stub in some upstream builds — pass keys positionally (dict_show exists for this).

• Serial desync (frame-error flood / connect timeout) happens if the process holding the port is killed mid-connection — unplug/replug the device; hw disconnect before reloading the plugin.

• A slot that starts as (disabled)undef must be hw slot enabled after loading or it won't emulate.

• The CLI tokenizes with a plain whitespace split — never quote arguments (hw slot nick ... -n bike, not -n "bike", or the quotes become part of the nick) and keep paths space-free.

Development

cd rango
uv run ruff check ai_shell/ plugin/mcp/          # lint
uv run python -m ai_shell.selfcheck            # cited commands exist in the real CLI tree,
                                               # gate semantics, simulated /plugins install boot
uv run python -m ai_shell.test_analyze         # dump-analyzer regression tests

CI runs the same three gates on every push and PR (see .github/workflows/ci.yml).

Layout: ai_shell/ wrapper library · plugin/ Kimi Code plugin (see plugin/README.md) · scripts/ upstream-CLI installer + plugin setup · adapters/ other-harness configs (see adapters/README.md) · AGENTS.md contributor/agent notes.

Credits

Everything device-side — the Chameleon Ultra firmware and the official CLI that Rango drives — comes from RfidResearchGroup/ChameleonUltra, © its authors (see its AUTHORS.md). Rango is only the AI copilot layer on top: it contains no upstream code, and the CLI is fetched by scripts/install-cli.sh and used unmodified.

Also standing on:

• the Proxmark3 community — known-key and MAD AID conventions (mad.json can be dropped into ~/.chameleon_ai/dicts/);

• NXP's MF1S50YYX datasheet — the access-bits and value-block decode tables in ai_shell/analyze.py follow it;

• the Model Context Protocol — the interface every supported agent harness speaks;

• Kimi Code — the plugin format and the default LLM endpoint for the standalone shell.

License & responsible use

Rango itself is MIT. The upstream Chameleon Ultra CLI is a separate project under GPL-3.0 — fetched, not vendored — so its license governs the CLI, not this repo. Operate only on cards and devices you own or are explicitly authorized to test.