# Give your agent production context across tools

Connect your existing agent to infrastructure, code and change evidence through Graph MCP. Compare it with querying each source directly.

Source: https://anyshift.io/for-agents/use-cases/connect-production-context

## Before starting

- Keep your existing AI agent. Reuse an Anyshift project with the relevant sources connected, or create one. Kubernetes collection requires an in-cluster Helm collector.

[Connect Graph MCP](https://anyshift.io/for-agents/product/context)

## Workflow

- Your agent needs to find which applications use a queue and which machines host them. The answer crosses application, Kubernetes and cloud records.

1. Connect Graph MCP through browser OAuth, or reuse the plugin that includes it.

2. Confirm the project, resolve the exact resource and inspect its recorded relationships and changes.

3. Return the dependency path with resource IDs, timestamps and any missing or truncated evidence.

## Set up

- Connect the sources you need: Use the integration catalog for permissions and collection setup. Do not add sources already connected to the intended project.

- Connect and make a first read: Use the Graph endpoint, authenticate and run the bounded first-read prompt. No API token belongs in the MCP configuration.

[Connect the sources you need](https://anyshift.io/for-agents/integrations)

[Connect and make a first read](https://anyshift.io/for-agents/product/context)

## Evidence

- Project scope: describe_schema and list_projects identify the available model and selected project. find_resources resolves names to stable hashedIDs.

- Coverage: get_related reports truncated or filtered relationship types. Check collection timestamps and the requested time window; a connected integration alone does not prove coverage.

[Project scope](https://docs.anyshift.io/pages/product/integration/graph_mcp)

[Coverage](https://docs.anyshift.io/pages/product/integration/graph_mcp#tools)

## Expected result

- A source-backed path from the queue to its applications and hosting machines. Unresolved connections stay unknown rather than being inferred from similar names.

- Three days of history are included by default; longer history depends on the plan. Verify the recorded timestamps and available evidence before attempting an older incident.

## 83% fewer model tokens in our cross-system benchmark

- Across ten tasks spanning AWS, Kubernetes, Datadog, GitHub and PagerDuty, prepared Graph queries used 83.1% fewer cumulative model tokens and took 46.7% less time on average than native MCPs.

- Anyshift’s September 2026 benchmark used one synthetic environment, the same model and time limit, and one selected pair per task, including incomplete answers. The Graph agent received prepared queries; native MCP agents chose their calls. This measures retrieval and answer generation, not installation or collection costs. Token reduction is not an equivalent reduction in spend.

[Read the benchmark and methodology (Markdown)](https://anyshift.io/markdown/blog/mcp-vs-graph-cross-domain-benchmark-v2.md)

## When are direct MCPs enough?

- Keep direct source access when one tool already has the answer, or when you need an operation outside the collected graph. Compare the Graph when repeated cross-system lookups, resource matching and change history slow your agent down.

## How should I compare them on my stack?

- Use the same question, model, permissions and time window. Check both answers against source records. Compare correct relationships, unknowns, tool calls, tokens and elapsed time. Count initial setup separately and repeat across several tasks; one fast answer does not establish better accuracy.

## Try a cross-system question

```text
Use Anyshift Graph MCP. Call describe_schema, then list_projects and confirm [project]. Resolve [queue] in [environment] to its hashedID. Find the applications using it and their hosting machines using only relationships exposed by this project's schema. Cite resource IDs, sources and timestamps when available. Report missing, ambiguous or truncated evidence; do not invent connections or change resources.
```

[Connection instructions](https://anyshift.io/for-agents/product/context)

[Graph tools](https://docs.anyshift.io/pages/product/integration/graph_mcp#tools)

## Usage

- Graph reads use the shared Graph query allowance. Your own AI provider bills model usage separately; Annie investigation runs use a separate allowance.

[Pricing](https://anyshift.io/for-agents/pricing)