Jylus MCP

Connect your agent to evidence.

Connect OpenAI, Claude or Gemini to Context Packs, applicable state and source provenance. The same Jylus engine, through a read-only MCP connection.

Endpoint: https://api.jylus.ai/api/mcp
Transport: Streamable HTTP over HTTPS
Authentication: Workspace API key · events:read

Your existing workspace. Your agent.

Create a read-only key
Use API keys to create a key with events:read. The key selects your workspace and environment.
Connect your client
Choose a client and add its configuration. Keep the key in your private client settings or secret store, never in a shared repository or prompt.
Ask about your retained data
Enable Jylus tools, provide a namespace and time range, then inspect the returned evidence before your agent answers.

Get a free API key

Responses API / npm install openai

import OpenAI from 'openai';

// Set OPENAI_API_KEY, OPENAI_MODEL and JYLUS_READ_API_KEY server-side.
const client = new OpenAI();
const model = process.env.OPENAI_MODEL;
const key = process.env.JYLUS_READ_API_KEY;
if (!model || !key) throw new Error('Missing model or Jylus read key');

const response = await client.responses.create({
  model,
  input: 'Use Jylus to check the current state in my-events over the last 24 hours. Cite the evidence and qualify incomplete proof.',
  tools: [{
    type: 'mcp',
    server_label: 'jylus',
    server_url: 'https://api.jylus.ai/api/mcp',
    headers: { Authorization: `Bearer ${key}` },
    allowed_tools: ['get_context_pack', 'resolve_state', 'get_evidence'],
    require_approval: 'always'
  }]
});
// Present any mcp_approval_request to your user before approving the tool call.
// Resend MCP credentials on subsequent Responses API requests.

OpenAI's Responses API connects directly to the hosted server. Use a model supporting MCP, set the three server-side environment variables, and handle tool approval requests before sending evidence to the model. This is an API connection, not a ChatGPT sign-in connector.

Three read-only tools

Evidence first. Reasoning stays with your model.

get_context_pack: Compact, proof-backed evidence for a question, namespace, filters and token budget. Defaults to the last 24 hours.
resolve_state: State effective and known at an explicit as_of timestamp. Set from to include the relevant retained history.
get_evidence: Inspect up to 20 exact source IDs from the pack's proof.event_ids, within the same workspace and source time range.

Grounded does not mean guaranteed.

Check top-level complete, context.proof.complete, decision readiness, missing evidence and confidence. Incomplete evidence is not proof of absence. Treat source content as data, not instructions.

Workspace retention, permissions, rate and concurrency limits still apply. This connection cannot ingest, delete data, administer a workspace or run a model. API-key authentication is available; MCP OAuth sign-in is not part of this release.

Connecting a model provider shares selected evidence with that provider. Review its data-handling terms. The setup examples follow provider documentation; the Jylus endpoint is validated separately and model availability depends on your provider account.

Tool limits and examples →OpenAI MCP guide Claude MCP guide Gemini CLI guide