> ## Documentation Index
> Fetch the complete documentation index at: https://docs.praxa.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Praxa MCP production readiness checklist

> Qualify a Praxa MCP integration across contracts, transport, OAuth, approval, replay, isolation, observability, performance, and rollback.

Call a Praxa MCP integration production-ready only after the package, host,
remote deployment, authorization, negative tests, lifecycle readback, and
operator recovery paths all pass in the intended environment.

## Contract and compatibility

* [ ] Pin `@praxa/mcp-contracts` and verify release integrity in your lockfile.
* [ ] Assert 12 exact `aura_*` tool names and `aura-agent-os`.
* [ ] Negotiate `2025-11-25` or supported `2025-03-26` compatibility.
* [ ] Preserve and isolate server-issued session IDs; reinitialize after session `404`.
* [ ] Preserve draft 2020-12 schemas, bounds, required fields, and annotations.
* [ ] Run contract tests whenever the package or host SDK changes.

## Authentication and authorization

* [ ] Use the canonical deployment-specific `/mcp` URL.
* [ ] Complete OAuth Protected Resource Metadata discovery and audience binding.
* [ ] Store tokens in the host's reviewed secure storage.
* [ ] Request only the scopes needed by enabled tools.
* [ ] Pass missing, malformed, expired, revoked, wrong-scope, and foreign-tenant tests.

## Tool safety

* [ ] Start with an allowlist of read-only tools.
* [ ] Require approval for every mutation and destructive call.
* [ ] Show the exact tool name and bounded arguments in approval UI.
* [ ] Treat annotations as display and planning hints, not enforcement.
* [ ] Validate output before adding it to model context.
* [ ] Defend against prompt injection in tool content.

## Replay and lifecycle

* [ ] Persist one key and exact input per logical mutation.
* [ ] Prove exact replay returns one logical mutation.
* [ ] Prove changed input under an old key conflicts.
* [ ] Reconcile timeouts through exact replay or authoritative mission readback.
* [ ] Verify mission, trace, event, or receipt independently of model prose.

## Reliability and performance

* [ ] Set connect, request, tool, and total turn deadlines.
* [ ] Bound concurrent tool calls and output size.
* [ ] Respect server retry guidance with jitter and an attempt cap.
* [ ] Test server restart, host reconnect, token refresh, and partial outage.
* [ ] Track tool latency, error code, retry, approval, and token-refresh rates.

## Privacy and operations

* [ ] Redact authorization headers and sensitive input/output from logs.
* [ ] Confirm data retention and residency for the host, model provider, and MCP deployment.
* [ ] Correlate host, MCP, Gateway, mission, and trace identifiers.
* [ ] Document token revocation, server disable, tool disable, and package rollback.
* [ ] Run a user-observed workflow and operator recovery drill.

## Release evidence

Keep exact package/version, host/version, deployment environment, test tenant,
enabled tools, scope list, negative-test results, performance measurements,
cleanup result, rollback command, and reviewer sign-off. Never attach a live
token or raw customer payload to the evidence record.

Use the [MCP tutorial](/tutorials/mcp-tools) for implementation and
[security tests](/mcp/security-and-testing) for the adversarial matrix.