Have your AI agents decide based on your values.

Two MCP tools that give any agent access to your personal values and validate its decisions against them. Works with Claude, ChatGPT, OpenClaw, and any MCP client.

Start the free assessment

See how it works ↓

If you build or run AI agents: Your agent makes decisions on your behalf ,what to recommend, what to send, what to schedule, what to skip. The model has no idea what you actually value. Generic guardrails catch policy violations; they don't catch "this is technically fine but not how I'd choose."

If you live in Claude or ChatGPT all day: The model has been telling you for two years to "define your core values" and add them to your project or custom GPT. But how? Listing five words off the top of your head isn't the same as actually knowing them and it certainly isn't enough for the model to weigh real trade-offs.

Krystallos is one MCP server that solves both. Spend 20–30 minutes once on a values discovery process refined over a decade with 380+ users. After that, any MCP-compatible AI, agentic or conversational, can read your values and validate its decisions against them.

Two tools your AI gets when you connect Krystallos — plus one utility helper.

get_values

Returns your personal values, grouped by conviction strength (Strong / Medium / Mild). No params. Use this to inject your values into any system prompt, project, or custom GPT.

  
    {
  "strong": ["Honest", "Curiosity", "True Friendship", "Self-Reliance"],
  "medium": ["Successful", "Independence", "Capable", "Health and Well-Being"],
  "mild":   ["Adventure", "Recognition", "Pleasure"]
}
  

validate_alignment

Pass an action your agent is about to take (and optional context). The server scores it against your full values profile, including the relationships and hierarchy you defined, and returns supporting values, conflicting values, and suggestions.

  
    Input:
{
  "action": "Accept the promotion that requires relocating",
  "context": "It's a 40% raise but I'd see family less"
}

Output:
{
  "alignment_score": 0.42,
  "supporting":  ["Successful", "Capable", "Independence"],
  "conflicting": ["True Friendship", "Stability", "Belonging"],
  "suggestions": [
    "Negotiate remote/hybrid before deciding",
    "Define a minimum visit cadence with family up front"
  ]
}
  

Use this anywhere your agent is about to commit to an action; pre-commit Slack messages, calendar holds, automated decisions, code commits, anything. Score below your threshold? Surface the conflict to the user before acting.

get_assessment_status

Utility tool. Lets your AI client detect whether you've finished the assessment yet (returns whether it's complete, which phase you're on, and how many values you have). Useful for handling first connections gracefully before you've done the assessment.

Honest engineering notes: get_values returns intensity grouping only. Your AI never sees your private hierarchy. validate_alignment uses the hierarchy server-side for scoring but never returns it. The action and context inputs are capped at 2,000 characters each (cost guard, not a soft limit).

Three steps. The first one's free and takes 20–30 minutes.

No account, no payment. Browser-based. You'll work through three short phases:

Rate all 76 values (Strong / Medium / Mild / Neutral / Negative) — about 10 minutes.
Cluster the ones that survive into groups that make sense to you — about 10 minutes.
Hierarchy for each group, decide which value is the parent — about 5–10 minutes.

No pre-made categories. No suggested groupings. The structure that comes out is yours. `

Works with

Confirmed working today:

Client	How to connect
Claude Desktop (Mac/Win)	New: Customize → Add plugins → Add MCP server (paste your regional URL). Older builds: Settings → Connectors.
Claude Code (CLI)	Drop the snippet (provided in-app after subscribing) into `.mcp.json`. Uses `mcp-remote`.
claude.ai (web)	Profile → Connectors → Add custom connector → paste your regional URL.
ChatGPT Desktop	Settings → Connected Apps → Add server → paste your regional URL. OpenAI's MCP support is still evolving so the process may shift.
OpenClaw	Settings → MCP servers → Add → paste your regional URL.
Any compliant MCP client	Krystallos implements MCP with OAuth 2.0 + PKCE and RFC 7591 Dynamic Client Registration. Any compliant client works.

Pick your data region

Your data lives in the AWS region you choose at signup. Use the URL for that region:

Region	MCP URL
United States	`https://mcp-us.apps.kaleidoscopeaxiom.com/mcp`
Canada	`https://mcp-ca.apps.kaleidoscopeaxiom.com/mcp`
European Union	`https://mcp-eu.apps.kaleidoscopeaxiom.com/mcp`

Use the URL that matches the region you signed up in. Pasting a different region's URL won't work, the regions are independent deployments. If you signed up in one region but want to switch later, contact us, moving values data between regions is supported, just not self-service yet.

Privacy and safety

Where your data lives

One AWS region. Your choice — United States, Canada, or European Union. Three independent regional deployments, each with its own user database, data tables, and secret store. No cross-region replication. Switching regions is a manual, auditable migration — not a background sync.

How your AI tool authenticates

OAuth 2.0 with PKCE. No API keys. No long-lived bearer tokens to leak.
Under the hood: your AI client self-registers using the open RFC 7591 standard; you authenticate via your Krystallos account; the server issues a 1-hour access token plus a 30-day refresh token. Cancel your subscription and connected clients lose access within ~65 minutes.

What get_values actually exposes

The intensity-grouped list (Strong / Medium / Mild) and nothing else. The relationships and hierarchy you defined during the assessment stay on the server. They're used internally by validate_alignment for richer scoring, but they're never returned to your AI client.

Where validate_alignment data flows

Your action and context strings go to an AWS Bedrock-hosted model for scoring, along with your full values profile. AWS Bedrock does not train its hosted models on customer input. That's an AWS contractual commitment, not a setting we configured.

What we never do

```
Sell your values data
```

Train any model on your values, actions, or alignment-validation results

Share your data across regions without an explicit migration request

Embed third-party tracking in the MCP server itself

Account deletion

One-click from your account settings. Email-verified. Deletes your values, your assessment history, your subscription record, and your account itself. GDPR-compliant by design.

Pricing

Two ways in

Plan	What you get
MCP only	All three MCP tools (get_values, validate_alignment, get_assessment_status). Connect any MCP-compatible AI client.
Mim (full coaching app)	Everything above, plus the Mim coaching app on iOS and Android: AI coaching with safety monitoring, journeys (career, health, relationships, growth), persistent memory across sessions.

The free Krystallos values assessment is and stays free. You only pay if you want your AI to use the results.

Pricing by region

Region	MCP only	Mim monthly	Mim annual
United States	$4.99 USD / mo	$14.99 USD / mo	$99.99 USD / yr (~$8.33/mo)
Canada	$6.49 CAD / mo + GST/HST	$19.99 CAD / mo + GST/HST	$129.99 CAD / yr + GST/HST (~$10.83/mo + tax)

Everywhere else Local currency via Stripe (web) or Apple / Google (Mim mobile). Payment processor handles currency conversion and any applicable taxes

Cancel anytime

Subscriptions are managed through your payment provider (Stripe on web, or Apple / Google for Mim mobile). Cancel from your account settings or your platform's subscription manager. Your connected AI clients lose access within ~65 minutes of cancellation. Your values data is retained until you delete your account.

Does Mim cover MCP access?

Yes. The Mim subscription unlocks the same MCP server with the same tools. There's no separate "MCP add-on" line item. One subscription, both surfaces.

Common questions

Yes. The MCP tools return a "complete your assessment first" message until you've finished. There's no way to seed values manually, the whole point is that the values come from the structured discovery process, not a list you typed in.
20–30 minutes for most users. Phase 1 (rating 76 values) is ~10 min. Phase 2 (clustering them into groups that make sense to you) is ~10 min. Phase 3 (deciding the parent value in each group) is ~5–10 min. You can pause and resume, your progress saves automatically, locally if you're not signed in, in the cloud if you are.
Typical 3-5 seconds end-to-end. get_values and get_assessment_status return in ~50–150 ms. Don't put validate_alignment in a tight inner loop. It's designed for "before you commit to an action," not "every token.”
Two ways:
• Quick tweaks: drag-and-drop directly in the Tree view to reparent values or adjust relationships in place.
• Bigger changes: retake the assessment.
Connected AI clients see the updated profile on the next call.
Connected AI clients lose access within ~65 minutes (your access token survives until it expires; the next refresh fails). Your values data is retained until you delete your account, so you can resubscribe later and pick up where you left off. Account deletion is one-click and permanent.
The MCP server is not open source today. The MCP standard itself is open (modelcontextprotocol.io), so any compliant AI client can talk to our server, but the server implementation, the values discovery process, and the alignment scoring logic are proprietary.
```
it's a useful signal, not a verdict. validate_alignment is a model reasoning over your values profile and the action you described, focused on a narrow task. We expose the score (0–1), the supporting and conflicting values, and suggestions, so you can decide for yourself whether the reasoning holds. Treat it as a second opinion, not an oracle.
Single user. One Krystallos account = one values profile = one MCP connection (across multiple AI clients, all sharing the same values). Shared or team values aren't currently planned, but we're always open to enhancement requests.
We'll give you 90 days' notice and a values export (the same data the MCP returns, plus your hierarchy). Your data is yours; we'd rather hand it back than disappear with it.

Ready to make your AI accountable to your values?

Start with the free Krystallos assessment. 20–30 minutes. No account, no payment. When your tree is built, hit Explore with AI and connect Claude, ChatGPT, OpenClaw, or any MCP-compatible client.