VibiumBLOCK
The verification layer for coding agents
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
[](https://www.npmjs.com/package/vibium) [](https://pypi.org/project/vibium/) [](https://central.sonatype.com/artifact/com.vibium/vibium) [](LICENSE)
The verification layer for coding agents.
Vibium gives AI agents the tools they need to check their work. Install the browser skill and your agent can navigate pages, fill forms, click buttons, and take screenshots — all through simple CLI commands. Also available as an MCP server and as JS/TS, Python, and Java client libraries.
New here? Get started in JavaScript, Python, or Java — zero to hello world in 5 minutes.
Why Vibium?
- AI-native. Install as a skill — your agent learns the full browser automation toolkit instantly.
- Zero config. One install, browser downloads automatically, visible by default.
- Standards-based. Built on WebDriver BiDi, not proprietary protocols controlled by large corporations.
- Lightweight. Single ~10MB binary. No runtime dependencies.
- Flexible. Use as a CLI skill, MCP server, or JS/Python/Java library.
Agent Setup
npm install -g vibium vibium setup
Agents can run vibium setup --non-interactive. Optional: install skills with npx skills add https://github.com/VibiumDev/vibium.
Select all Vibium skills in the installer, then choose your agent and installation scope. vibium setup installs Chrome when it is missing.
- browser — explore pages, automate actions, record sessions, and carry out browser tasks from plain-language instructions.
- check — independently assess whether a claim holds in the browser or a recording.
213997b734d5OBSERVED · 2026-10-20Install
Commands as the repository documents them. They are shown, not run.
git clone https://github.com/VibiumDev/vibium.git
npx skills add .
pip install -e . # Editable install - code changes take effect immediately
pip install vibium
claude mcp add vibium -- vibium mcp
npm install -g vibium
Host compatibility
What the documentation claims. We have not run a compatibility test.
| Host | Status | Notes |
|---|---|---|
| claude-code | mentioned | |
| claude-desktop | mentioned | |
| codex | mentioned | |
| cursor | mentioned | |
| gemini-cli | mentioned |
What it tells the agent
The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.
--- name: browser description: Automate browsers with the Vibium CLI. Use to navigate websites, inspect pages, fill forms, extract page data, debug UI behavior, capture screenshots and recordings, or delegate browser goals with vibium run. --- # Vibium Browser Automation — CLI Reference The `vibium` CLI automates Chrome (and Firefox, via `--engine firefox`) from the command line. The browser auto-launches on first use (daemon mode keeps it running between commands). Use this skill for browser automation, exploration, debugging, and recording. Choose explicit CLI commands when you know the steps, or use `vibium run` when a model should work out how to accomplish a browser goal. For an independent acceptance verdict, use the `check` skill if installed, or `vibium check "<claim>"`. Run’s `completed` result and your own browser observations do not substitute for that invocation. ## Core Workflow For direct browser commands, use this pattern: 1. **Navigate**: `vibium go <url>` 2. **Map**: `vibium map` (get element refs like `@e1`, `@e2`) 3. **Interact**: Use refs to click, fill, select — e.g. `vibium click @e1` 4. **Re-map**: After navigation or DOM changes, get fresh refs with `vibium map` ## Binary Resolution Before running any commands, resolve the `vibium` binary path once: 1. Try `vibium` directly (works if globally installed via `npm install -g vibium`) 2. Fall back to `./clicker/bin/vibium` (dev environment, in project root) 3. Fall back to `./node_modules/.bin/vibium` (local npm install) Run `vibium --help` (or the resolved path) to confirm. Use the resolved path for all subsequent commands. **Windows note:** Use forward slashes in paths (e.g. `./clicker/bin/vibium.exe`) and quote paths containing spaces. ## Browser readiness During initial setup, run `vibium ready browser --json` with the engine/channel that the workflow will use. It inspects installed browser and driver files without launching them or touching existing sessions. Passing confirms the installation, not browser launch or BiDi connectivity. If installation is missing, use the reported `vibium install` command and retry. Direct browser work does not need AI configuration. Do not rerun readiness before every action when setup is unchanged. ## Delegate a browser goal with Run Use `vibium run "<goal>"` when the task is clear but the sequence of browser actions needs investigation. For known steps, use the commands below directly. Run uses the existing local Chrome or Firefox session and a fresh model context. Keep the same `--session` or `VIBIUM_SESSION` throughout the workflow. Load the project's configured AI settings in the shell running the CLI. Settings in an environment file must use exported assignments (`export NAME=value`). Run `vibium ready ai --json` during initial setup or after configuration changes; `result.ready: true` means the provider tool round-trip passed. Never print credentials. Direct browser commands do not require a model or AI readiness. With a settings page already open: ```sh vibium run "Change the timezone to America/Chicago and save it" --json -o browser-run.zip ``` `vibium "<multiword goal>"` is shorthand for Run. Use explicit `run` in scripts or when the prompt could be mistaken for a command. Read `result.status` (`completed` or `not_completed`), the summary, and evidence. Execution failures return an error. If independent verification is needed, follow with the `check` skill or `vibium check "<claim>"` in the same session; Check starts another fresh model context. `-o` saves a recording to a new path. An existing recording is exported without stopping it. A browser that Run starts closes afterward unless `--keep-open` is set; a browser already open stays open. Run can change application state. Run and Check share `VIBIUM_AI_*` defaults. Per-call `--provider`, `--model`, `--ai-base-url`, and `--reasoning-effort` also work with `vibium ready ai`. When changing provider, supply a model explicitly; inherited endpoint and effort settings are cleared. Credentials remain in the provider's environment variable. Use the project's chosen provider rather than silently switching it. ## Command Chaining Chain commands with `&&` to run them sequentially. The chain stops on first error: ```sh vibium go https://example.com && vibium map && vibium click @e3 && vibium diff map ``` **When to chain:** Use `&&` for sequences that should happen back-to-back (navigate → interact → verify). Run commands separately when you need to inspect output between steps. **When NOT to chain:** Don't chain commands that depend on parsing the previous output (e.g. reading map output to decide what to click). Run those separately so you can analyze the result first. ## Commands ### Discovery - `vibium map` — map interactive elements with @refs (recommended before interacting) - `vibium map --selector "nav"` — scope map to elements within a CSS subtree - `vibium diff map` — compare current vs last map (see what changed) ### Navigation - `vibium go <url>` — go to a page - `vibium back` — go back in history - `vibium forward` — go forward in history - `vibium reload` — reload the current page - `vibium url` — print current URL - `vibium title` — print page title ### Reading Content - `vibium text` — get all page text - `vibium text "<selector>"` — get text of a specific element - `vibium html` — get page HTML (use `--outer` for outerHTML) - `vibium find "<selector>"` — find element, return `@e1` ref (clickable with `vibium click @e1`) - `vibium find "<selector>" --all` — find all matching elements → `@e1`, `@e2`, ... (`--limit N`) - `vibium find text "Sign In"` — find element by text content → `@e1` - `vibium find label "Email"` — find input by label → `@e1` - `vibium find placeholder "Search"` — find by placeholder → `@e1` - `vibium find testid "submit-btn"` — find by data-testid → `@e1` - `vibium find xpath "//div[@class]"` — find by XPath → `@e1` - `vibium find alt "Logo"` — find by alt attribute
Trust audit
BLOCKgrade F · trust 40/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | WARN |
| L1 | Static analysis of the code | FAIL |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | PASS |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- declared (9 observation(s))
- Network
- declared (8 observation(s))
- Shell
- declared (7 observation(s))
- Dependencies
- not all pinned
- Secrets in source
- none-found
Findings (25)
({ stdout } = await exec(binary, [...flags, ...args], { env, timeout: 230000, maxBuffer: 4 * 1024 * 1024 }));await exec(binary, ['check', '--help'], { env, timeout: 5000 });try { await exec(binary, ['daemon', 'stop'], { env, timeout: 15000 }); }gradle-wrapper.jar
var-parts-trace.zip
mod = __import__(mod_name, fromlist=[cls_name])
cmd.Example += example + ` --provider local --model my-model --ai-base-url http://127.0.0.1:8080/v1 --reasoning-effort ""` + "\n # Uses these settings for one invocation; credentials still come from
baseURL = "http://127.0.0.1:8080/v1"
return "http://127.0.0.1:8080/v1"
const pTrace = pTraceFile ? parseNDJSON(join(playwrightDir, pTraceFile)) : null;
if (vTrace && pTrace) {const pTypes = groupBy(pTrace, (e) => e.type);
const pCtxOpts = pTrace.find((e) => e.type === "context-options");
const pActions = pTrace.filter((e) => e.type === "before" || e.type === "after");
first-check-var-parts.webm
recording-sample.webm
broken.webm
.gitmodules
try { ({stdout}=await exec(bin,['--headless','--json',...args],{env,timeout:240000,maxBuffer:16*1024*1024})); }await exec(bin,['daemon','stop'],{env,timeout:15000}).catch(()=>{});cd ../../../clients/python && pip wheel . -w dist --no-deps
pip install -e ../../packages/python/$(PYTHON_PLATFORM_PKG) -e ".[test]"; \
$(TIMEOUT_CMD_ABS) python -m pytest ../../tests/py/ -v --tb=short -x -n $(PY_PARALLEL) --dist=loadfile \
--ignore=../../tests/py/test_browser_modes.py \
--ignore=../../tests/py/test_firefox.py
Gates applied: no_behavioural_pass.
213997b734d5full audit observations/trust-audit/skill/vibiumdev__vibe-check.json · Report an issue / request a re-scanAudit history
Every audit this skill has had.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-20 | 213997b734d5 | BLOCK | F | 40 | first audit |
Questions
What does the vibe-check — vibiumdev/vibium skill do?
The verification layer for coding agents
Is vibe-check — vibiumdev/vibium safe to install?
No — not without reading the findings first. The audit graded it F (40/100) and found 3 critical or high issues in the source. Each one is listed on this page with the file and line it is on.
What can vibe-check — vibiumdev/vibium access on my machine?
The audit observed that it reaches the network, runs shell commands and reads or writes files. Each of those is consistent with what it says it does. Secrets in the source: none found.
Which assistants does vibe-check — vibiumdev/vibium work with?
Its documentation mentions claude-code, claude-desktop, codex, cursor and gemini-cli. That is what the text claims, not a compatibility test we ran.
How current is this page?
The grade is for one exact copy of the source (213997b734d5), read on 2026-10-20. The repository is watched, and a new audit runs when it changes — this is the first audit.
Add this badge to your README
Shows this page's audit result and links back to it. It updates when the audit does.
[](https://modelsatlas.com/skills/vibiumdev/vibe-check/)
Not your listing, or something out of date? Claim or update your listing.