> For clean Markdown of this page, append .md to its URL. For the complete documentation index, see https://www.superagent.sh/llms.txt.


Shared security memory for agents, PR scanners, and red-team assessments through MCP.

# Interfaces

Security Graph gives security agents a shared memory of what has already been investigated. It connects assets, reports, findings, evidence, comments, review decisions, vulnerability types, and compliance mappings. Later assessments can retrieve that context rather than start from an empty history.

For Superagent users, connect to the hosted MCP endpoint with your organization API key. Superagent operates the graph service and database. You can also self-host an independent Security Graph with your own PostgreSQL database, agents, and models.

## Choose your setup

| Setup | Connection | Database and credentials |
| --- | --- | --- |
| [With Superagent](https://www.superagent.sh/docs/security-graph/superagent) | Connect your agent to the hosted `https://www.superagent.sh/mcp`; Superagent scanners receive graph tools automatically | Superagent manages storage; connect with your organization API key |
| [Bring your own database](https://www.superagent.sh/docs/security-graph/bring-your-own-database) | Run Security Graph on your infrastructure and connect agents to its MCP endpoint | Your PostgreSQL database, deployment-issued keys, and access rules |

Both use the same graph engine and classification catalog. A standalone database is separate from your hosted Superagent organization; it does not appear in the hosted dashboard automatically.

## Read, investigate, record

1. **Read earlier context.** Find the target asset, retrieve relevant findings, and inspect their evidence, comments, and review history.
2. **Investigate the current target.** Check whether the earlier evidence and assumptions still apply to the current code or deployment.
3. **Record the result.** Save a structured finding, select vulnerability types and compliance mappings, and retain the supporting rationale.
4. **Review and reuse.** Comments and status changes become context for the next assessment.

This helps agents recognize previously investigated behavior, accepted risks, and false positives. Historical context is evidence to reassess; a prior dismissal is not an instruction to suppress every similar issue.

## What the graph connects

| Record | Context it provides |
| --- | --- |
| Assets | Repositories, applications, agents, packages, and other assessment targets |
| Reports | The assessment that produced a finding and its execution or review state |
| Findings | Impact, reproduction evidence, policy assessment, and remediation |
| Comments and status history | Who supplied review context, what changed, and why |
| Vulnerability types | The technical weakness selected by the assessing agent |
| Compliance entries | The framework or taxonomy references selected as relevant to the finding |

`get_context` follows these connections with a maximum depth of three and up to 100 returned records. Check its truncation flag. Use `get_neighbors` for paged connections and filtered list tools to narrow larger datasets.

## Agent-selected classifications

Security Graph provides two independently selectable lists: `vulnerability_types` and `compliance`. The agent evaluates the evidence and selects from both. The service validates and saves those selections; it does not run Julia-1 or another classification model.

Search the current catalog instead of copying a fixed vocabulary into your prompts:

```json
{
  "intent": "Find classification choices for confirmed code injection.",
  "query": "code_injection",
  "limit": 25
}
```

Pass these arguments to `get_classification_options`. Refine the query or follow `next_cursor` with the same filters and limit. Copy the returned `catalog_version` into the classification write, together with exact `type_ids`, `compliance_ids`, and an evidence-based `rationale`. Both arrays are required; empty arrays explicitly mean that no choice applies.

The catalog includes CWE, OWASP Web/API/LLM, MITRE ATLAS, ASVS, NIST frameworks, and additional compliance references. Coverage varies: some collections contain individual controls, while ISO 27001, SOC 2, PCI DSS, and AIUC-1 have more limited reference coverage. Read the returned framework coverage metadata before interpreting a mapping. A mapping expresses relevance, not certification or proof of a control failure.

The MCP service publishes and validates the catalog. Superagent manages catalog updates for hosted users, and connected agents discover the available choices through MCP. Stored selections retain their catalog version so historical labels do not change meaning.

## Tool contracts

Every graph call requires an `intent`: a short explanation of the security task, without secrets. Writes require an `idempotency_key`; reuse the same key and exact payload when retrying the same operation. Updates also use the current version returned by a read, so concurrent edits produce a conflict rather than silently overwrite one another.

Discover the input schemas from your connected MCP server. Native Superagent scanners have additional run-scoped tools; these are not general-purpose tools exposed to organization API keys.

## Next steps

- [Use Security Graph with Superagent](https://www.superagent.sh/docs/security-graph/superagent)
- [Bring your own database](https://www.superagent.sh/docs/security-graph/bring-your-own-database)
- [Superagent MCP reference](https://www.superagent.sh/docs/mcp#security-graph-tools)
- [Security Graph source and deployment guides](https://github.com/superagent-ai/security-graph)

---
Source: https://www.superagent.sh/docs/security-graph
Index: https://www.superagent.sh/llms.txt