Your filter scored 100%. Then a goblin in a fake moustache walked in. Here's the free tool that catches him.
Everything you need, in one place
A free gift from Enki Labs: no account, no sign-up, no catch. MIT licensed, so it's yours to keep, change and share.
1 · Download & install
Windows (one click)
- Download enki-harness-0.1.1-windows.zip (10 MB).
- Unzip it anywhere and double-click Install enki-harness.cmd.
- It brings its own Python, copies itself to
%LOCALAPPDATA%\EnkiHarness, finds your AI tools and asks before connecting each one. - Claude Desktop asks which folder it may test filters in: press Enter to use
Documents\enki-harness-projects. - Restart your AI tool. enki-harness is now in its tools.
Mac / Linux
- Download enki-harness-0.1.0.zip (80 KB) and unzip it.
- With Python 3.11 (nothing else needed):
python -m hb install
Windows may warn you the first time because the scripts aren't code-signed yet. To remove it: Uninstall enki-harness.cmd or python -m hb uninstall.
2 · What you get
- Twin cases → catches over-blocking: every must-catch case gets a near-identical innocent twin that must pass.
- Rules switched off (alone and in pairs) → finds dead rules and rules that cancel each other out.
- An attacker slice → leetspeak, dots, stretched spellings: the evasions you didn't think to write.
- Controls that must fail → if always-allow or always-block passes, the harness tells you it's broken, not your filter.
- A falsifier, timestamped → you say what proves you wrong before the run, so the goalposts can't move.
- Bugs first, score last → every miss with the exact message and the rule to blame. A 0% or 100% rate is flagged "check the instrument".
- Case ideas & test data → suggested terms and hard negatives, plus pointers to vetted public datasets.
- Inside your AI tool → Claude Code, Claude Desktop, Cursor, VS Code (Copilot), Codex and Windsurf, over MCP.
- Private by design → runs on your machine: no account, no telemetry, no network calls.
3 · How to use it
In your AI tool: just ask:
you ▸ test my filter in filters/swear.py with enki-harnessIn your browser: python -m hb serve, open http://127.0.0.1:8765, describe what the filter must catch and must let through, press run. A working example is already filled in.
Your filter is a Python module with RULES (name → check) and classify(text, enabled_rules):
# filters/swear.py
RULES = {
"hate_word": lambda t: "hate" in t.lower(),
"kill_word": lambda t: "kill" in t.lower(),
}
def classify(text, enabled_rules):
return any(RULES[n](text) for n in enabled_rules)Reading the report: gates first (all must be ✓), then every bug row with the rule to blame. Fix the rule it names, run again, repeat. The score only matters once the rows are clean.
For sexual content, minors, self-harm and hate speech it never writes examples, only points to vetted datasets.
you ▸ test the filter in examples/planted_filter.py with enki-harness GATES ✓ prereg falsifier recorded and hashed before the run ✓ arms all arms ran ✓ control always-allow and always-block both failed, as required ✓ independence attacker, twins and your cases come from different authors BUGS # every row the full system gets wrong, with the rule to blame ✗ Two rules cancel each other out quote_exception × game_exception "She said "kill the boss in the game" to me" ✗ Dead rule: removing it changes nothing loser_caps ✗ Flags an innocent twin "I whatever you" blame: hate_word ✗ Missed a must-catch case "you are so dumb" ✗ Attacker evasions "1 h4t3 y0u" "I h.a.t.e y.o.u" "I haaate yooou" +12 SCORE # last, after the rows falsifier verdict: FALSIFIED
Real output from the bundled example: a filter with four bugs planted on purpose.
For every case it must catch, a near-identical innocent one it must not flag.
Dumb stand-ins (always-allow, always-block) that must fail, or the harness itself is broken.
Each rule switched off, alone and in pairs, to find dead rules and rules that cancel each other out.
An independent slice written by a different generator than your cases: leet, dots, stretched spellings.
You say what result proves you wrong before it runs. It's timestamped, so the goalposts can't move.
No score until every gate is green, and the score comes last. A 0% or 100% rate is flagged "check the instrument".
- Windows: download enki-harness-0.1.1-windows.zip, unzip it and double-click Install enki-harness.cmd. It brings its own Python, finds your AI tools and asks before connecting each one. Restart the tool and enki-harness is in it.
- Mac / Linux: download enki-harness-0.1.0.zip (standard-library Python 3.11, nothing else to install), then
python -m hb install. Orpython -m hb serveand openhttp://127.0.0.1:8765: a working example is filled in. Press run. - Point it at your own filter: a Python module exposing
RULESandclassify(text, enabled_rules). That's the whole contract.
On Windows the installer does this for you. Elsewhere, one command registers it as an MCP server in the tools it finds on your machine: Claude Code, Claude Desktop, Cursor, VS Code with GitHub Copilot, OpenAI Codex (and Windsurf). It asks before touching each one, backs up every config first, and changes only its own entry.
$ python -m hb install # then just ask your agent: you ▸ test my filter in filters/swear.py with enki-harness
It only runs filters inside the project folder your tool opens.
Uninstall enki-harness.cmd (Windows) or python -m hb uninstall removes it again.
Runs on your machine. No account, no telemetry, no network calls. Your filter and your cases never leave it.
For sexual content, minors, self-harm and hate speech it never generates examples, only pointers to vetted public datasets.
The local page accepts only its own requests, and every input has a hard limit, enforced on the server.
MIT licensed. Built by Enki Labs, makers of Enki, a companion that remembers you. Questions: [email protected]