LizzyDocs
Docs navigation MCP server

Agent integration

Connect an agent to the MCP server

Connect Claude Code, Cursor, or the Anthropic API to Lizzy’s durable distillation tools with a full-access API token.

Connect an MCP client

Lizzy exposes a stateless MCP Streamable HTTP server at https://lizzy.albinilabs.com/v1/mcp. Send every JSON-RPC request with POST andAuthorization: Bearer <token>. The server returnsapplication/json; it does not open an SSE stream or issue anMcp-Session-Id.

POST/v1/mcp

Supports initialize, notifications/initialized,ping, tools/list, and tools/call with protocol versions 2025-03-26 and 2025-06-18.

Create an API token with the full_access scope athttps://lizzy.albinilabs.com/tokens. The workspace must also have Pilot enabled. Start with an lz_test_ token: its tools read and write test-mode data end to end. Use a separate lz_live_ token only when the client should act on live data.

bashClaude Code
claude mcp add --transport http lizzy https://lizzy.albinilabs.com/v1/mcp   --header "Authorization: Bearer lz_test_your_token_here"
json~/.cursor/mcp.json
{
  "mcpServers": {
    "lizzy": {
      "url": "https://lizzy.albinilabs.com/v1/mcp",
      "headers": {
        "Authorization": "Bearer lz_test_your_token_here"
      }
    }
  }
}
bashAnthropic Messages API MCP connector
curl https://api.anthropic.com/v1/messages   -H "Content-Type: application/json"   -H "X-Api-Key: $ANTHROPIC_API_KEY"   -H "anthropic-version: 2023-06-01"   -H "anthropic-beta: mcp-client-2025-11-20"   --data '{
    "model": "claude-opus-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Inspect my Lizzy workspace."}],
    "mcp_servers": [{
      "type": "url",
      "url": "https://lizzy.albinilabs.com/v1/mcp",
      "name": "lizzy",
      "authorization_token": "lz_test_your_token_here"
    }],
    "tools": [{"type": "mcp_toolset", "mcp_server_name": "lizzy"}]
  }'
bashinitialize with curl
curl --fail-with-body --silent --show-error -X POST   "https://lizzy.albinilabs.com/v1/mcp"   -H "Authorization: Bearer lz_test_your_token_here"   -H "Content-Type: application/json"   --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}'
bashlist tools with curl
curl --fail-with-body --silent --show-error -X POST   "https://lizzy.albinilabs.com/v1/mcp"   -H "Authorization: Bearer lz_test_your_token_here"   -H "Content-Type: application/json"   -H "MCP-Protocol-Version: 2025-06-18"   --data '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}'

Use the ten declared tools

Start workflow work with distill_workflow. Its actions start, inspect, resume, switch, or abandon one durable workflow. The six phase tools are valid only in their matching phase.

distill_workflowStart, inspect, resume, switch, or abandon a workflow.
distill_dataConfigure or inspect the workflow's data.
distill_rewardsChoose and inspect reward signals.
distill_planPrepare and validate the training plan.
distill_trainingRequest or inspect training.
distill_resultsRead run evidence, reports, and produced artifacts.
distill_deploymentRequest or inspect a deployment transition.
workspace_stateRead a bounded snapshot of loops, datasets, runs, reports, deployments, and balance.
product_docsSearch these product docs.
api_referenceRead the policy-tagged REST catalog.

Follow the saved workflow phase

Distillation is one durable workflow with seven phases. Workflows are shared per workspace, so every MCP client and the Lizzy dashboard see the same saved workflows. Inspect and continue an existing workflow instead of recreating it in another client.

textworkflow phases
scope → data → rewards → plan → training → results → deployment

Every tool result uses the same envelope. next_actions names the tool or tools that may continue the recorded state.

texttool result envelope
{status, workflow_id, phase, summary, refs, blockers, next_actions}

Treat the token as full authority

The same policy layer as the Pilot copilot classifies every underlying write. The MCP agent still acts with the full authority of the supplied full_access API token. Connecting the token authorizes the agent to perform every action that the MCP tools expose in that token's test or live mode.

Account for service limits

Rate limit120 MCP tool calls per minute per workspace. Inspect the mcp_calls bucket in GET /v1/rate_limits.
Cold startThe first call after about 30 minutes idle can take about 10 seconds.
Supported clientsAPI tokens work with Claude Code, Cursor, and the Anthropic API MCP connector. Claude.ai web connectors are not yet supported because they require OAuth.
BillingMCP tool calls are not billed. Live managed teacher calls are billed per token at catalog rates, with margin included. Test-mode calls record cost but are not charged. Runs and other triggered actions use their normal billing rules.