# Anyshift for agents

Anyshift connects infrastructure, code and observability data in a production graph. Query that graph with your own agent, or ask Annie to investigate an incident.

Canonical page: https://anyshift.io/for-agents

## Choose an interface

### Graph MCP

Read topology, dependencies and change events with your own agent.

Endpoint: `https://api.anyshift.io/mcp/graph`

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

#### Connect with Codex

Reuse an existing connection. For a new connection in Codex, run these commands, sign in through browser OAuth, then confirm the intended project with list_projects.

```bash
codex mcp add anyshift-graph --url https://api.anyshift.io/mcp/graph
codex mcp login anyshift-graph
```

#### Other MCP clients

Your client must support Streamable HTTP and browser OAuth. Add this remote server, sign in to Anyshift and select the intended project. The configuration format varies by client; reuse an existing plugin or server connection.

```json
{
  "mcpServers": {
    "anyshift-graph": {
      "type": "http",
      "url": "https://api.anyshift.io/mcp/graph"
    }
  }
}
```

#### Make a first read

```text
Use the existing Anyshift Graph MCP connection. Call describe_schema, then list_projects and confirm the bound project is the one I intend to query. If the project cannot be confirmed, stop before reading resources. Choose a non-secret resource kind from the schema and use find_resources to list at most 5 resources. Return their names, kinds, scopes and hashedIDs. If none match, report the empty result; do not invent resources or switch projects.
```

A successful read returns resource identities from the intended project. An empty result confirms that the request completed, not that your infrastructure is empty.

### Annie Remote MCP

Delegate an investigation to Annie and retrieve its analysis and next steps.

Endpoint: `https://api.anyshift.io/mcp/remote`

[Connect Annie](https://docs.anyshift.io/pages/product/integration/remote_mcp)

#### Connect with Codex

Reuse an existing connection. For a new connection in Codex, run these commands, sign in through browser OAuth, then confirm the intended project with list_projects.

```bash
codex mcp add annie --url https://api.anyshift.io/mcp/remote
codex mcp login annie
```

#### Other MCP clients

Your client must support Streamable HTTP and browser OAuth. Add this remote server, sign in to Anyshift and select the intended project. The configuration format varies by client; reuse an existing plugin or server connection.

```json
{
  "mcpServers": {
    "annie": {
      "type": "http",
      "url": "https://api.anyshift.io/mcp/remote"
    }
  }
}
```

#### Ask Annie to investigate

```text
In the selected Anyshift project, investigate this alert in read-only mode: [paste the alert, affected service and time window]. Use ask_annie with mode=rca, then get_annie_response until the investigation completes or reports a failure. Return the findings, supporting evidence and unresolved questions. Do not change infrastructure or publish messages.
```

Annie returns findings and evidence, or explains what could not be established. A started investigation is not a completed result.

## When does the Graph help?

- Use Graph MCP when your agent repeatedly joins resources and changes across infrastructure, code and monitoring. Start with direct source MCPs when one tool already contains the evidence your task needs.
- Anyshift’s cross-system benchmark measured 83.1% fewer model tokens and 46.7% less time with prepared Graph queries. It covered ten tasks in one synthetic environment; it does not measure setup costs or guarantee savings on other tasks.

[Compare Graph with direct MCPs](https://anyshift.io/for-agents/use-cases/connect-production-context) · [Benchmark and methodology (Markdown)](https://anyshift.io/markdown/blog/mcp-vs-graph-cross-domain-benchmark-v2.md)

## Start with a task

- Add production context to an existing coding or incident agent through Graph MCP.
- Review a pull request with your own coding agent using Graph MCP. The agent needs the diff; the graph supplies production context.
- Investigate PagerDuty alerts automatically and send findings to a configured Slack channel, or start an on-demand run through Annie Remote MCP.

[Review Kubernetes permissions before merge](https://anyshift.io/for-agents/use-cases/review-kubernetes-permissions) · [Investigate PagerDuty alerts. Get findings in Slack.](https://anyshift.io/for-agents/use-cases/investigate-deployment-alert) · [Give your agent production context across tools](https://anyshift.io/for-agents/use-cases/connect-production-context)

## Connect to your project

- Reuse your existing Anyshift integrations and MCP connection. If you already installed the plugin, Graph MCP is included; do not register it twice.
- For a new connection, follow the setup guide for your client, sign in through browser OAuth and select your project. Both endpoints use Streamable HTTP; no API token belongs in the MCP configuration.
- No Anyshift project yet? Create one and connect the data sources needed for your task. An existing project only needs additional integrations when the required evidence is missing.

[Install the plugin](https://docs.anyshift.io/pages/product/integration/production_intelligence_agent_plugin) · [SDK](https://docs.anyshift.io/pages/product/integration/sdk) · [Graph API](https://docs.anyshift.io/pages/product/graph-api/overview) · [CLI](https://docs.anyshift.io/pages/product/integration/cli) · [Create a project](https://app.anyshift.io/sign-up) · [Project setup](https://docs.anyshift.io/pages/onboarding/quick_start)

## Work with graph evidence

- Call describe_schema first to discover the selected project's resource types, relationships and events.
- Use find_resources to resolve a resource, then use its hashedID to inspect relationships and time-bounded changes. Names may repeat across clusters and namespaces.
- Graph MCP is read-only. Inspect returned timestamps, page limits and get_related completeness signals. A truncated or filtered result is not a complete dependency map; missing evidence is not proof of absence.
- The graph keeps three days of history by default; longer history depends on your plan. Check that the incident window is actually covered. Current health cannot establish past health, and related events alone do not establish a root cause.
- Use Cypher only with Graph MCP's query_graph tool. The Graph SDK, API and CLI have their own Graph Query Language; queries are not interchangeable.

[Graph tools and examples](https://docs.anyshift.io/pages/product/integration/graph_mcp#tools) · [Graph Query Language](https://docs.anyshift.io/pages/product/integration/graph_query_language) · [Data history](https://docs.anyshift.io/pages/overview)

## Let Annie investigate

- Annie runs incident investigations. Use its web app, Slack or Remote MCP, or configure PagerDuty triggers.
- For PR review and deployment preparation, connect your own agent to Graph MCP. Your agent performs the analysis; Anyshift provides the production evidence.

[Explore incident investigation](https://anyshift.io/for-agents/product/remote-agents)

## Queries and runs are different

- A Graph query is a request for data from your Anyshift graph. Use the plugin, Graph MCP, SDK, API or CLI: all share the same query allowance.
- Anyshift agents use a separate run allowance. When you use your own AI provider to query the graph, you pay that provider separately. See pricing for current plan limits.

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

## If you get stuck

### Sign-in or project access fails

Re-authenticate the existing connection through your client. Use list_projects to check available projects; only use set_project for the project you intend to access. If the project remains unavailable or the returned scope is inconsistent, stop and contact Anyshift support.

[Authentication help](https://docs.anyshift.io/pages/product/integration/graph_mcp#connect)

### No matching resources or recent evidence

Check the selected project, resource filters and time window. Inspect the existing integration's permissions and collection status before adding another source. Missing or stale evidence does not establish health or recovery.

[Integration help](https://docs.anyshift.io/pages/onboarding/overview_integrations)

### Annie is still working

Poll the existing investigation with get_annie_response. Do not start another investigation just because the first call has not returned a final answer.

[Investigation help](https://docs.anyshift.io/pages/product/integration/remote_mcp#first-use)

## Official references

- [Documentation index](https://docs.anyshift.io/llms.txt)

- [Graph API specification](https://raw.githubusercontent.com/anyshift-io/anyshift-graph-sdk/main/openapi/graph-api.v1.json)

- [Security and access](https://docs.anyshift.io/pages/privacy_security/faq)

- [Quickstart](https://docs.anyshift.io/pages/onboarding/quick_start)