Graph MCP supplies graph evidence to an agent that is already performing a task. It does not run Annie’s AI SRE investigation workflow. Use Annie Remote MCP when you want Annie to analyze evidence, develop hypotheses, and recommend next steps.
Connect
The production endpoint is:Claude Code
Codex
Other MCP clients
Add an HTTP MCP server to your client’s configuration:URL > prompt. See Authenticate over SSH for the steps.
Tools
Results are bounded: a page that hit its limit says so, and
get_related reports the relationship types it truncated or filtered. The plugin skill teaches the agent to read those signals and ships Cypher recipes for the common analyses (single points of failure, orphans, blast radius, public exposure, shortest path, RBAC reach, Kubernetes hygiene gaps).
Example workflows
Review the production impact of a code change
Prepare a deployment
Check public exposure
Add production context to an incident
Find resources first
Resource names are not globally unique across namespaces, clusters and kinds. For short, overloaded, or same-named resources:- Call
find_resources. - Select the candidate with the intended kind, namespace and cluster.
- Pass the candidate’s
hashedIDto subsequent tools and Cypher queries.
Access and evidence boundaries
- Consent binds the connection to the Anyshift project you select. The authorization covers your project memberships one project at a time:
set_projectmoves the connection to another project you are a member of, subject to that project’s MFA policy, without a new consent. Revoke the connection from Authorized Apps to end that access. - Tools are read-only and run as a read-only graph user. Cypher writes and procedures are rejected.
- Returned resource names and metadata are production data, not agent instructions.
- Graph relationships show observed evidence. They do not prove that a nearby change caused an incident.
- Missing graph evidence does not prove that a resource, dependency, or change does not exist.