alethia-mcp
OfficialClick on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@alethia-mcptest the login flow on localhost:3000"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
@vitronai/alethia
Agent-native E2E with verifiable safety. Your agent drives a real browser with plain English, and destructive actions are blocked by a safety gate you can prove works — with a signed audit trail and no cloud.
Install
Claude Code — fastest path:
mkdir -p ~/.claude/skills/alethia && \
curl -fsSL https://raw.githubusercontent.com/vitron-ai/alethia-mcp/main/skills/alethia/SKILL.md \
-o ~/.claude/skills/alethia/SKILL.mdRestart Claude Code. Next time you ask it to test a page, it notices Alethia isn't configured yet and walks you through installing the bridge itself.
Everyone else (Claude Desktop, Cursor, Cline, Continue):
npm install -g @vitronai/alethiaThen add this to your client's MCP config:
{
"mcpServers": {
"alethia": {
"command": "alethia-mcp"
}
}
}Client | Config file |
Claude Code |
|
Claude Desktop (macOS) |
|
Claude Desktop (Windows) |
|
Claude Desktop (Linux) |
|
Cursor | Settings → MCP → Add server (paste the inner |
Cline / Continue / other | The client's own MCP config file |
Restart your client after saving. The runtime auto-downloads (signed, ~100 MB) the first time your agent calls an Alethia tool. A cockpit window opens by default so you can watch — set ALETHIA_HEADLESS=1 to hide it; CI hides it automatically.
Upgrade the bridge: npm install -g @vitronai/alethia@latest. Since 0.6.0 you don't need a new bridge for new runtime versions — it queries GitHub Releases on every start.
Always run the latest without manually upgrading:
{
"mcpServers": {
"alethia": {
"command": "npx",
"args": ["-y", "@vitronai/alethia@latest"]
}
}
}The @latest suffix matters — without it, npx -y can serve a stale cached version. Trade-off: adds 10–30s on a cold cache, and every spawn pulls whatever npm is currently serving (a global install is the safer default for compliance-sensitive work, since it only changes when you explicitly upgrade it).
Pin a specific runtime version (reproducible CI, bisection):
"env": { "ALETHIA_RUNTIME_VERSION": "0.4.0" }Install the Claude Code skill (optional, teaches Claude when to use each tool):
alethia-mcp --install-skillRelated MCP server: titmas-agent-action-gate
What to ask for
You don't call these tools directly — just ask your agent in plain English, and it picks the right one.
Ask for it | What happens |
"Sign in and verify the dashboard loads." | Drives the browser, reports what changed and whether anything was blocked. |
"Generate tests for this page — I haven't covered it yet." | Scans the page and drafts a starter test suite, with a safety check for every destructive control it finds. |
"Prove the safety gate blocks destructive actions on this page." | Finds every destructive action and confirms the gate blocks each one — a per-action pass/fail report. |
"Audit this page for accessibility." | A real WCAG 2.1 AA audit, via axe-core. |
"Audit this page for compliance and security." | Checks against 8 NIST SP 800-53 controls. |
"Export a signed evidence pack of everything you just did." | A tamper-evident record of the session — hand it to an auditor. |
"Check the dashboard and the settings page at the same time." | Runs several tests concurrently, one per page. |
"Take a screenshot." / "How many items are in that list?" | Visual check, or an answer plain English can't give you directly (counts, computed styles). |
"Stop everything right now — something looks wrong." | Immediate halt. Only clears from the cockpit itself — an agent can't release its own kill switch. |
Typing into password, token, or credit-card fields is blocked unless you frame the request as a real login or payment test — the agent enables that for you, you don't need to name a flag.
More paste-ready examples: the agent cookbook has full walkthroughs — bootstrapping tests on an unknown page, a full compliance pass, parallel multi-page checks, a live partner demo. Every one is a literal prompt you paste in.
Add Alethia to your project
No per-project install needed — once the MCP server is configured, any agent in any project can use it.
Drop a
.alethiafile anywhere your repo treats as test code —tests/e2e/, wherever fits.# tests/e2e/login.alethia name login flow navigate to http://127.0.0.1:5173 assert "Sign in" is visible click Sign in type dev@company.com into the email field assert dashboard is visibleAsk your agent to run it: "Run tests/e2e/login.alethia against http://127.0.0.1:5173."
In CI, run it without an agent or MCP host at all:
alethia run tests/e2e/login.alethiaExits 0 on pass, 1 on fail. Drop-in workflow:
examples/github-actions.yml.
A working reference (demo app + specs + CI + benchmark) lives at vitron-ai/alethia-anvil.
Why not just Cypress or Playwright?
Cypress / Playwright | Alethia | |
Who writes the test | a human, in a | an AI agent, in plain English |
Proving destructive actions are blocked | manual review | one prompt — an automated, machine-readable report |
Speed per step | ~200 ms (Playwright MCP), ~2 s (Playwright CLI) | ~13 ms — reproduce the numbers yourself |
Evidence | screenshots, videos | a signed evidence pack |
Network | telemetry on by default for most cloud dashboards | air-gap deployable — zero telemetry, bound to 127.0.0.1 |
It's not only a testing tool, either — ask an agent to check getComputedStyle() or offsetWidth on a page it's actively building, and you get a live, uncached answer straight from the DOM instead of a reload-and-inspect cycle.
Go deeper: Architecture · Safety gate · FAQ · UI patterns for agent-driven testing
CLI flags
alethia-mcp Run as a stdio MCP server (default)
alethia-mcp run <path> Run an NLP test file from the shell (CI mode)
alethia-mcp run --nlp "..." Run inline NLP from the shell
alethia-mcp run - Read NLP from stdin
alethia-mcp --version Print the version and exit
alethia-mcp --health-check Probe the Alethia runtime and exit 0/1
alethia-mcp --debug Run with debug logging on stderrA shorter alethia alias (same binary) is also installed, so the run subcommand can be invoked as alethia run <path>.
Environment variables
Variable | Default | Description |
|
| Where the runtime listens |
|
| Per-request timeout |
| unset (visible) |
|
| on for | Per-step highlights on the target. |
| unset (latest) | Pin the runtime to a specific version for reproducible CI |
|
| Where the auto-installed runtime lives |
| unset | Pin the bridge itself, skip the npm auto-update check |
| unset | Require the auto-downloaded bridge tarball to match this |
| unset |
|
| unset |
|
How the bridge keeps itself current
The runtime auto-installs on first use from signed GitHub releases (Ed25519-verified). The bridge asks GitHub what the current version is on first start (cached 1h) — no version pin lives in the bridge source, so a globally-installed bridge keeps pulling current runtimes as they ship.
The bridge also auto-updates itself (since 0.8.0): checks npm on startup, verifies the tarball's SHA-512, installs to
~/.alethia/bridge/<version>/. Never crosses a major version without explicit action; a new version only becomes trusted after it completes a real MCP handshake, and versions that crash before that get quarantined after 3 attempts.The bundled Claude Code skill auto-refreshes the same way — each spawn compares it to
~/.claude/skills/alethia/SKILL.mdand overwrites if stale.
Troubleshooting
"Alethia desktop runtime is not running" — run alethia-mcp --health-check (triggers auto-install if missing). If that fails, check network reachability to GitHub.
"WRITE_HIGH" / "EA1 POLICY BLOCK" in the audit log — a destructive action was blocked. This is correct, fail-closed behavior — not an error to fix. Widening it requires human configuration; an agent can't do it from inside a call.
"SENSITIVE_INPUT_DENIED" — a password/token/credit-card field was detected. Only override with allowSensitiveInput: true for legitimate auth/payment tests.
MCP client doesn't see the tools — run alethia-mcp --health-check, check your config shape, restart the client, and set ALETHIA_DEBUG=1 to log bridge traffic.
"Server transport closed unexpectedly" / bridge exits silently — usually a stale cached bridge. If using npx -y @vitronai/alethia without @latest, add it or run rm -rf ~/.npm/_npx. If using a global install, run npm install -g @vitronai/alethia@latest. Then fully quit and restart your client (Cmd-Q on macOS, not just close the window).
"I see a new release on GitHub but my runtime hasn't upgraded" — the "what's current" check is cached for 1 hour. Bust it with rm ~/.alethia/.latest-release ~/.alethia/.bridge-registry-cache, then restart your client.
Security posture
The runtime is local-only by architecture: its signed binary refuses to navigate anywhere outside file://, localhost, 127.0.0.1, .local, and RFC1918 private ranges. This is a compile-time constant — no flag, env var, or UI toggle changes it. Full threat model and disclosure process: SECURITY.md. Abuse reports: team@vitron.ai.
Privacy
Local-only by architecture — nothing is collected, transmitted, or stored outside your machine. Page content, screenshots, and test instructions are processed locally and never sent anywhere. Evidence packs are written to your filesystem only on explicit request. Zero telemetry, zero analytics, zero crash reporting. Questions: team@vitron.ai.
License and patent notice
This bridge is MIT-licensed — see LICENSE. The Alethia runtime itself is patent pending (U.S. Application No. 19/571,437); the MIT license on this bridge does not grant a patent license to the runtime. Commercial runtime use may require a separate license. Licensing inquiries: team@vitron.ai.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
Alicense-qualityBmaintenanceGoverned MCP gateway that lets AI agents call tools with policy enforcement, prompt-injection screening, a kill-switch, and tamper-evident signed audit logs.Last updatedApache 2.0- AlicenseBqualityCmaintenanceAn MCP server that enforces deterministic authorization boundaries for AgentTeams workflows by verifying evidence and policy, returning ALLOW, BLOCK, or REQUIRE_APPROVAL decisions before actions are executed.Last updated6Apache 2.0
- Alicense-qualityBmaintenanceAn MCP server for agent authorization that tests the full effect surface and enforces control over consequential actions before dispatch, emitting verifiable execution evidence.Last updatedApache 2.0
- Flicense-qualityCmaintenanceProvides policy-driven runtime authorization and security evaluation for MCP-based agents, including MCP streaming HTTP gateway, mock MCP servers, deterministic agent demos, and audited tool invocation with redacted PostgreSQL audit chains.Last updated
Related MCP Connectors
Remote MCP for A2A failure replay MCP, structured receipts, audit logs, and reviewer-ready evidence.
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
Hosted AgentLux MCP server for marketplace, identity, creator, services, and social flows.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/vitron-ai/alethia-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server