@myth-os/mcp

MCP server that gives AI agents persistent identity context, governance boundaries, quality eval, and cross-session memory. Connects to a mythOS backend via HTTP.

Your agent forgets who it's working for every session. This fixes that.

What it solves

• Identity amnesia -- Agents start every session blank. mythos_getContext loads the creator's voice, values, and venture context so the agent knows who it's serving.
• Ungoverned execution -- Agents do whatever you prompt them to. mythos_guard returns hard constraints and narrative boundaries before the agent acts.
• No eval feedback loop -- Agents can't tell if their output quality is degrading. mythos_evaluate returns pass rates, coherence scores, and diagnostics.
• Zero cross-session memory -- Agents don't remember what happened yesterday. mythos_search and mythos_recap give full memory retrieval across sessions.
• Confabulation without grounding -- Agents invent context. mythos_consult runs 12 independent governance perspectives against your intent before execution.
• Approval gate -- Agents execute without sign-off. mythos_approve lets the creator approve or reject a governed plan before execution begins.

Tools

| Tool | Description | |------|-------------| | mythos_getContext | Load creator identity, voice, values, and venture context. Call at session start. | | mythos_guard | Get active constraints, execution limits, and narrative boundaries for a venture. | | mythos_evaluate | Check quality health of recent work. Returns pass rates, angel/demon scores, diagnostics. | | mythos_search | Search creator memory -- facts, preferences, relationships, prior decisions. | | mythos_recap | Get recent session history -- decisions, milestones, context from previous work. | | mythos_consult | Run a full governance consultation. 12 archetypal seats evaluate your intent and produce an execution plan. | | mythos_approve | Approve or reject a consultation plan before execution begins. |

Install

npm install @myth-os/mcp

Configure

Create ~/.mythos/config.json:

{
  "apiKey": "your-api-key",
  "deploymentUrl": "https://your-deployment.convex.site",
  "userId": "your-user-id"
}

Or use environment variables:

export MYTHOS_API_KEY="your-api-key"
export MYTHOS_DEPLOYMENT_URL="https://your-deployment.convex.site"
export MYTHOS_USER_ID="your-user-id"
export MYTHOS_WORLD_ID="optional-venture-id"

Usage with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "mythos": {
      "command": "npx",
      "args": ["mythos-mcp"],
      "env": {
        "MYTHOS_API_KEY": "your-api-key",
        "MYTHOS_DEPLOYMENT_URL": "https://your-deployment.convex.site"
      }
    }
  }
}

Usage with Claude Code

Add to your .mcp.json:

{
  "mcpServers": {
    "mythos": {
      "command": "npx",
      "args": ["mythos-mcp"],
      "env": {
        "MYTHOS_API_KEY": "your-api-key",
        "MYTHOS_DEPLOYMENT_URL": "https://your-deployment.convex.site"
      }
    }
  }
}

Usage with any MCP client

The server uses stdio transport. Run it as a subprocess:

npx mythos-mcp

It speaks the Model Context Protocol standard. Any MCP-compatible client works.

Tool details

mythos_getContext

Returns the creator's narrative identity -- who they are, how they communicate, what they value. Optionally scoped to a specific venture/world.

Parameters:
  worldId? (string) -- Scope to a specific venture. Omit for creator-level context.

mythos_guard

Returns governance constraints -- what the agent is allowed to do, execution limits, narrative boundaries, and active policies.

Parameters:
  worldId? (string) -- Scope to a specific venture.

mythos_evaluate

Returns quality health metrics for a time window. Includes pass rates, angel/demon scoring (optimism vs skepticism), deterministic checks, and diagnostic messages.

Parameters:
  windowMinutes? (number) -- Lookback window. Default: 1440 (24 hours).

mythos_search

Searches the creator's persistent memory. Returns memory cards with confidence scores, scoped to actor-level and world-level knowledge.

Parameters:
  limit?   (number) -- Max results. Default: 10.
  worldId? (string) -- Scope to a specific venture.

mythos_recap

Returns recent session observations -- what happened across previous agent sessions. Gives cross-session continuity.

Parameters:
  limit?   (number) -- Number of recent entries. Default: 5.
  worldId? (string) -- Scope to a specific venture.

mythos_consult

Runs a full governance octave -- 12 archetypal seats independently evaluate your intent, surface tensions, and produce a governed execution plan. Use before significant implementation work. Supports pause/resume when the system needs more context.

Parameters:
  intent?   (string) -- What you're planning to do.
  octaveId? (string) -- Resume a paused consultation (instead of intent).
  inputs?   (object) -- Context to fill gaps at pause point.
  worldId?  (string) -- Scope to a specific venture.

mythos_approve

Approves or rejects a consultation plan that's waiting for creator review.

Parameters:
  octaveId (string)  -- The consultation to approve or reject.
  approved (boolean) -- true to proceed, false to reject.

Architecture

The MCP server is a thin stdio-transport wrapper around the mythOS HTTP API. No Convex SDK dependency at runtime -- just fetch(). The backend runs on Convex (serverless, reactive).

Agent <--MCP/stdio--> mythos-mcp <--HTTP--> mythOS Convex backend

Requirements

• Node.js >= 18
• A mythOS backend deployment (self-hosted or managed)
• API key with appropriate permissions

License

MIT