Agentic MCP, catalog discovery, reviewed graph memory, durable verified tasks, source mode, workflows, watches, and OAuth. agentic agentic agentic agentic/_index.md

Agentic

Agentic

MCP, catalog discovery, reviewed graph memory, durable verified tasks, source mode, workflows, watches, and OAuth.

Agentic GraphJin gives AI clients a discoverable, policy-aware surface for data access and controlled operations. Use the built-in agent or connect an AI client over MCP, then choose the smallest path that matches the request.

01 · Ask

Get an answer now.

GraphJin discovers the relevant graph, checks policy, validates the request, runs it, and returns the evidence behind the answer.

Use the server-side agent
02 · Prove

Know the agent still works.

Build verified tasks from your catalog, check both answers and methods, and gate model, prompt, or GraphJin changes against a trusted baseline.

Evaluate the agent
03 · Continue

Carry declared intent across runs.

Owner-scoped tasks warm-start later agent runs, journal their provenance, stay visible through notices, and can require a saved query to prove the declared outcome before close.

Use durable verified tasks
04 · Watch

Keep a standing question running.

A cursor-backed watch records changes. An optional tool-free AI flow can summarize, score, or suppress noise before the right conversation wakes.

Choose a watch and flow
05 · Act

Run only the action that was approved.

Workflows can use GraphJin tools, but autonomous watch delivery remains paused until the exact current action hash is confirmed.

Understand workflows

Approved catalog annotations add reviewed organizational context to exact catalog detail and semantic recall without changing source metadata. They remain untrusted data, carry author and approval attribution, and become stale historical notes when their addressed catalog entity vanishes.

The lead agent holds conversations, the caller’s skills, and their memories, and asks a stateless data worker the questions. Its threads, skills, and memories are GraphQL rows, so a GUI can manage them without a REST client.

Mission Control is the human view over the same state: task trails, unseen watch events, and observed annotation drafts live in one caller-scoped console page, while Agent chat turns response notices into links and task-pin actions.

The intended loop is catalog first, security second, validation or preview third, and then an answer or action. That sequence keeps agents useful without handing them raw database credentials or arbitrary shell access. The fastest way to see the full surface is the coffee-roastery demo.

Server-Side Agent Let GraphJin run the catalog-first discovery loop for you and return a typed answer through one MCP tool and one REST endpoint. Evaluate the DeepORG Agent Build a verified suite from your catalog, catch wrong answers and unsafe methods, and gate agent changes against a trusted baseline. Lead Agent The lead agent holds the conversation, your skills and your memories, and asks a stateless data worker the questions. Every part of it is readable and writable through GraphQL. MCP Connect AI clients to GraphJin through catalog-first Model Context Protocol tools. Catalog Graph Let agents discover databases, tables, columns, relationships, operations, capabilities, and examples. Security Graph Expose effective policy and high-risk findings before an agent acts. Source Mode Use source-local configuration, capabilities, and reloads for agentic deployments. Artifacts Overlay Use gj_artifacts for caller-scoped saved queries, fragments, workflows, and addressed catalog annotations. Declared Tasks Persist an owner-scoped goal, keep it visible across agent runs, and optionally require a declared saved-query check before close. Catalog Annotations Attach durable organizational notes to catalog entities, review them explicitly, and make approved context discoverable without changing source metadata. Watches Cursor-backed standing questions with optional inline, tool-free AxFlow triage before durable inbox, webhook, workflow, or MCP delivery. Choosing Watches, Flows, and Workflows Choose notification, AI triage, and autonomous action from two simple questions, then review everything through gj_watch. Mission Control Review declared task trails, unseen watch events, and annotation drafts from the same map used by GraphJin agents. Workflows Run named, policy-controlled workflow steps that can call GraphJin tools and GraphQL. MCP OAuth Protect hosted MCP endpoints with resource metadata, DCR/CIMD, PKCE, and audience checks.
Docs