Skip to content

Integrate

MCP server

Let a compatible AI coding agent use Reliastra: 18 tools over stdio or Streamable HTTP, authenticated with a scoped API key.

What the MCP server is

The MCP server exposes Reliastra over the Model Context Protocol: capability discovery, the public observatory, incident and evidence retrieval, workspace reads, dependency analysis, integration planning, and two narrowly scoped writes. It is an HTTP client of the canonical API and nothing else - no database, no duplicated business logic. What it can do is exactly what the API and the caller credential permit.

Transports

TransportHow to reach itAuthentication
stdioRun python -m mcp_server --transport stdio from a backend checkout with RELIASTRA_API_KEY set. For a local agent that spawns the process.API key from the environment; without one only public tools work
Streamable HTTPhttps://api.reliastra.com/mcp. For remote callers.Always required: Authorization: Bearer <rel_...> per request. No anonymous access exists on this transport.

Legacy HTTP/SSE is deliberately not offered. Only the official Python SDK client has been exercised against the server (protocol 2025-11-25 negotiated); other MCP clients speak the same standard surface but have not been tested, so verify before depending on one.

Install and configure

  1. 01

    Get a Reliastra API key

    Create one with reliastra keys create <name> [--scopes <list>], or in the console under /settings/developer. Reads need read scopes; the two write tools need write scope. The organization is pinned to the key: cross-org access is impossible by construction.

  2. 02

    Connect a local agent over stdio

    Check out the repository, install the server dependencies, and point the agent at the process. No package is published: the server ships with the backend.

    bash
    pip install -r backend/requirements-mcp.txt
    RELIASTRA_API_KEY=rel_... python -m mcp_server --transport stdio
    python -m mcp_server --check-config   # validate configuration, exit
  3. 03

    Or connect remotely

    Point any MCP-compatible client at https://api.reliastra.com/mcp with the key as a Bearer token. A copy-paste client configuration for both transports is on the MCP feature page.

  4. 04

    Verify with one call

    reliastra_get_capabilities needs no credential and returns the permission-aware capability document. If it answers, the transport, auth and server are all working.

Tools

18 tools. Read-only unless marked (write). API key required where marked (key). Every tool returns a uniform {ok, data, meta} envelope with its source endpoint and request id; failures are typed errors, never empty successes. The list below is generated from the same catalog the feature page renders, and a test diffs both against the server registration.

`reliastra_get_capabilities`
Discovery — The verified capability registry, filtered to your credential.
`reliastra_search_observatory`
Observatory — Search listed vendors by name or category.
`reliastra_get_provider`
Observatory — One vendor: detail, endpoints, recent status.
`reliastra_get_provider_metrics`
Observatory — Per-window availability and latency aggregates.
`reliastra_get_provider_timeline`
Observatory — Observation time series for a window.
`reliastra_get_pipeline_health`
Observatory — Whether the measurement pipeline itself is healthy.
`reliastra_search_incidents`
Incidents and evidence — Filterable search over observed vendor incidents.
`reliastra_get_public_incident`
Incidents and evidence — One incident: detail plus its evidence descriptor.
`reliastra_get_incident_evidence`
Incidents and evidence — The frozen evidence document for a public incident.
`reliastra_verify_evidence`
Incidents and evidence — Independent verdict on a record. Counts only when verified is exactly true.
`reliastra_list_dependencies` (key)
Workspace — Monitored dependencies on this account.
`reliastra_get_dependency_observations` (key)
Workspace — Recent observations for one dependency.
`reliastra_list_workspace_incidents` (key)
Workspace — Incidents on this account.
`reliastra_get_evidence_report` (key)
Workspace — One evidence report with its verification id.
`reliastra_analyze_dependencies`
Analysis and planning — Match an inventory against the catalog. Returns gaps — or that Reliastra adds no value.
`reliastra_plan_integration`
Analysis and planning — A concrete REST, CLI, or GitHub Action plan where justified, or no-use-case where not.
`reliastra_register_dependency` (key) (write)
Writes — Monitor a new HTTP endpoint. Never accepts stored headers.
`reliastra_trigger_check` (key) (write)
Writes — One immediate diagnostic check. Not a second scheduler.

Resources and prompts

  • Resources: reliastra://capabilities, reliastra://methodology, reliastra://limitations, reliastra://integration-guide, and reliastra://observatory/{provider} - machine-readable product truth, no tool call needed to read them.
  • Prompts: reliastra_workflow_assess_observability, reliastra_workflow_investigate_failure, reliastra_workflow_evaluate_fit, reliastra_workflow_plan_integration - reusable workflows that guide tool use without forcing a recommendation.

What the evidence establishes

Tools return what the API recorded and nothing more: timestamps, provider identity, observation scope, provenance, methodology version, and stated limits. Vendor-level evidence never proves a specific application request failed, and correlation is never presented as causation. Scores below 50 are often unknown, which is a valid result.

Troubleshooting

SymptomMeaning
401 invalid_tokenMissing, malformed, or non-rel_ credential. Send Authorization: Bearer rel_....
Tool error forbiddenThe credential was rejected upstream or lacks the scope. Check the key is active and the plan includes API access.
502 on the remote endpointThe MCP process is not running on the host. Operator-side; nothing the caller can fix.
Public tool works, workspace tool fails closedNo credential on the call. Provide the API key and retry - this is the designed behavior, not an outage.

Source and full operator documentation live with the implementation in backend/mcp_server/ (start at its README), including architecture, threat model, and the 427-test suite that guards it.

Problem with this page?

These guides live in the repository. An inaccuracy in the documentation is a bug, and the fastest fix is an issue with the sentence quoted. Open an issue.

Next

Methodology →

The canonical explanation of every rule and number RELIASTRA reports, and the limits of each.

Read the methodology, or run the thing.

Every rule in these guides has a page stating its thresholds and its limits. The other way to check it is to add a dependency and watch what the detector does.