MCP Server

Quick setup

Add the hosted Hebbrix endpoint to your client's MCP server configuration, save, and reconnect. Guest access does not require an account, API key, local process, or package installation.

{
  "mcpServers": {
    "hebbrix": {
      "url": "https://mcp.hebbrix.com/mcp"
    }
  }
}

Connection details

What the MCP server exposes

The tool surface covers memory, temporal graph, grounded reasoning, procedures, account workflows, and learning from explicitly reported outcomes. Representative tools include:

• hebbrix_remember stores durable memory from an agent interaction.
• hebbrix_ask answers using memory and temporal graph evidence.
• hebbrix_graph_query queries entities, relationships, timelines, and current truth.
• hebbrix_graph_status reports whether asynchronous graph enrichment is processing, failed, ready, or temporarily unreadable. Readiness is based on durable projection acknowledgements. Its related-memory count is not an entity-relationship count: zero related memories does not mean there are no extracted edges.
• hebbrix_choose_action records a decision receipt and chooses from measured outcome evidence.

Grounding and safety contract

Retrieval tools return only evidence that the API marked as grounded. Each verified response includes the grounding state, evidence memory IDs, and a claim receipt. If the upstream response is malformed, missing its safety envelope, degraded, or ungrounded, the hosted adapter fails closed instead of presenting related text as an answer.

When GraphRAG synthesis abstains but authoritative search has verified evidence,hebbrix_ask returns it with synthesis_status: "retrieval_only". It does not silently discard recall or misrepresent retrieved text as model synthesis. Failed tool calls use MCP isError=true consistently over hosted HTTP and local stdio.

MCP responses expose request IDs, backend builds, grounding versions, and cache-path diagnostics. Ask includes separate reasoning and retrieval diagnostics when fallback is needed. Failure categories distinguish unknown facts from unavailable verification, service failures, quota limits, and malformed evidence receipts. These diagnostics contain no credentials or memory content; they do not bypass the evidence checks. The companion API reports claim-grounding-v36.

Verify the connection

1. Reconnect your client and confirm that the Hebbrix tools appear.
2. Ask the agent to remember a distinctive, non-sensitive fact.
3. Start a fresh conversation and ask the agent to retrieve that fact.
4. For graph validation, store two related facts and query their relationship or timeline.

Guest and account-scoped use

Guest access is the quickest way to evaluate the hosted MCP connection. Applications that need account-owned collections, explicit tenant scoping, API administration, billing, or production data controls should use an authenticated Hebbrix account and follow the authentication guide.

Coding-agent workflows

For code-change history, previous-error lookup, fix tracking, and the dedicated/v1/coding-agent/* API surface, continue to the coding-agent guide.