Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.
Python agentic-aiai-securityalert-triageblue-teamcybersecuritydetection-engineeringincident-responselanggraphllm-agentsmcpmitre-attackmodel-context-protocolpurple-teamsecurity-operationsself-hostedsiemsoarsocthreat-detectionthreat-intelligence
-- 069: which third-party MCP servers a tenant has chosen to trust, and how far.
--
-- Gap-closure Phase 5.2.
--
-- An MCP server is somebody else's code, reached over the network, whose
-- replies land in a prompt that decides what an agent does next. That makes
-- this table a trust boundary rather than a list of endpoints, and every
-- column below exists because leaving it out would have widened that boundary
-- by default.
--
-- Why the allowlist is a column and not a setting
-- -----------------------------------------------
-- `tool_allowlist` defaults to the empty array, so a server saved with no
-- further thought advertises nothing to the model. The alternative default,
-- "everything the server offers", puts the choice of what the agent may call
-- in the hands of the server operator rather than the tenant, and a server
-- that later adds a tool would silently widen the agent's reach with no
-- change on this side. The allowlist is checked before dispatch, in the
-- agents service, so a tool outside it never reaches the network.
--
-- Why a timeout and a byte cap are stored rather than inferred
-- ------------------------------------------------------------
-- Both are the tenant's answer to "how much of my investigation budget may
-- this third party consume". A global default would be one operator's answer
-- imposed on every deployment, and an unbounded reply is a denial of service
-- against the agent loop and a way to push a payload past whatever scanning
-- is applied. They are `NOT NULL` with CHECK bounds so no row can carry
-- "unbounded" at all.
--
-- Why stdio has its own columns, and why it is off by default
-- ----------------------------------------------------------
-- A stdio MCP server is a local process the agent starts. That is a different
-- risk from an HTTP request: it is code execution on the agents container,
-- and the argument vector is attacker-influenced the moment anything derives
-- it from tenant data. `transport` defaults to `streamable_http`, and the
-- agents service refuses a stdio row outright unless an operator has both
-- enabled stdio and allowlisted the command. The columns exist so the refusal
-- can name what was configured instead of failing on a shape it cannot parse.
--
-- Why the credential is a vault token and not a column
-- ----------------------------------------------------
-- `auth_config` holds `CredentialVault` ciphertext (`vault:v1:` / `vault:v2:`)
-- under the same convention connector credentials already use, rather than a
-- bespoke `token` column. One decryption path, one rotation procedure, one
-- place a reviewer has to look.
CREATE TABLE IF NOT EXISTS aisoc_mcp_servers (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
tenant_id UUID NOT NULL REFERENCES tenants(id) ON DELETE CASCADE,
-- The namespace segment in `mcp.<name>.<tool>`. Constrained to the shape
-- an OpenAI function name may take, because it becomes part of one: a
-- name outside this set would produce a tool the provider rejects, and
-- the failure would surface as "the model never called it".
name TEXT NOT NULL CHECK (name ~ '^[a-z0-9][a-z0-9_-]{0,38}[a-z0-9]$'),
label TEXT,
transport TEXT NOT NULL DEFAULT 'streamable_http'
CHECK (transport IN ('streamable_http', 'stdio')),
url TEXT,
command TEXT,
args JSONB NOT NULL DEFAULT '[]'::jsonb,
-- CredentialVault ciphertext, same convention as connector auth_config.
auth_config JSONB NOT NULL DEFAULT '{}'::jsonb,
-- Empty means the agent may call nothing. Deliberate: see the header.
tool_allowlist JSONB NOT NULL DEFAULT '[]'::jsonb,
timeout_seconds INTEGER NOT NULL DEFAULT 20
CHECK (timeout_seconds BETWEEN 1 AND 120),
max_response_bytes INTEGER NOT NULL DEFAULT 65536
CHECK (max_response_bytes BETWEEN 1024 AND 1048576),
-- A registered server is not a reachable one. Registration and use are
-- separate decisions so an operator can write the row down, read what it
-- discovers, and only then let an investigation reach it.
enabled BOOLEAN NOT NULL DEFAULT FALSE,
created_by UUID REFERENCES users(id) ON DELETE SET NULL,
created_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
updated_at TIMESTAMPTZ NOT NULL DEFAULT NOW(),
UNIQUE (tenant_id, name),
-- Three properties, and the third is conditional on purpose.
--
-- A row may never carry both a URL and a command, and its target may
-- never contradict its declared transport: those are states nothing can
-- recover from, because there is no way to tell which field the operator
-- meant.
--
-- A row that names *neither* is only wrong once it is enabled. Disabled
-- is the default and is the half-finished draft an operator saves while
-- they go and find the URL; refusing that would make the console's own
-- create-then-configure flow impossible. An **enabled** row naming no
-- target is a server the agent would try to reach and cannot, so that is
-- refused here. The API refuses it earlier with a 422 naming the field,
-- and the agents service refuses it again at discovery with a sentence an
-- operator can act on; this is the floor under both.
CONSTRAINT aisoc_mcp_servers_transport_target CHECK (
(url IS NULL OR command IS NULL)
AND (transport <> 'streamable_http' OR command IS NULL)
AND (transport <> 'stdio' OR url IS NULL)
AND (NOT enabled OR COALESCE(url, command) IS NOT NULL)
)
);
COMMENT ON TABLE aisoc_mcp_servers IS
'Third-party MCP servers one tenant has chosen to trust, with the bounds '
'on that trust: an explicit tool allowlist, a timeout and a response-size '
'cap. Empty allowlist means the agent may call nothing.';
COMMENT ON COLUMN aisoc_mcp_servers.tool_allowlist IS
'Tool names the agent may call on this server. Checked before dispatch in '
'the agents service, so a tool outside it never reaches the network. '
'Defaults to empty, which advertises nothing.';
COMMENT ON COLUMN aisoc_mcp_servers.auth_config IS
'CredentialVault ciphertext under the same convention connector '
'credentials use. Never returned to a console caller.';
COMMENT ON COLUMN aisoc_mcp_servers.transport IS
'streamable_http | stdio. stdio starts a local process on the agents '
'container and is refused there unless an operator has enabled it and '
'allowlisted the command.';
CREATE INDEX IF NOT EXISTS idx_mcp_servers_tenant
ON aisoc_mcp_servers (tenant_id, enabled);
-- Row-level security. The `OR current_tenant_id() IS NULL` arm is what lets
-- the cross-tenant workers (retention purge, tenant deletion) reach this
-- table; dropping it makes those silently see nothing. FORCE so the table
-- owner, which is the role that runs migrations, does not walk past it.
ALTER TABLE aisoc_mcp_servers ENABLE ROW LEVEL SECURITY;
ALTER TABLE aisoc_mcp_servers FORCE ROW LEVEL SECURITY;
DROP POLICY IF EXISTS aisoc_mcp_servers_tenant ON aisoc_mcp_servers;
CREATE POLICY aisoc_mcp_servers_tenant ON aisoc_mcp_servers
USING (tenant_id = current_tenant_id() OR current_tenant_id() IS NULL);
-- 061 set ALTER DEFAULT PRIVILEGES so new tables pick these up, but only for
-- tables created by the role that ran it. Granting explicitly means a chain
-- replayed by a different owner still leaves the runtime role able to work.
DO $$
BEGIN
IF EXISTS (SELECT 1 FROM pg_roles WHERE rolname = 'aisoc_app') THEN
GRANT SELECT, INSERT, UPDATE, DELETE ON aisoc_mcp_servers TO aisoc_app;
END IF;
END
$$;