mcp

Temporal Cortex MCP is open scheduling infrastructure for AI agents, with the goal of giving any AI agent reliable and autonomous scheduling capabilities. Whether or not the other party in a conversation uses an AI agent, uses Google Calendar or Outlook, and whether replies are immediate or delayed by several days, scheduling can still be completed. The project's current version is v0.9.1 (March 2026), written in Rust, and distributed via npm and Docker.

[Purpose] It addresses common failures of AI agents in calendar tasks: even the latest LLMs score below 50% on temporal reasoning tasks, and earlier models scored as low as 29% on scheduling tasks and 13% on duration calculations. Common problems include choosing the wrong "next Tuesday," misjudging time zones, and double-booking calendars. Most calendar tools for AI are just thin CRUD wrappers and do not have temporal awareness, conflict detection, or safety fallbacks. Temporal Cortex provides temporal awareness, atomic booking, computed availability, deterministic RRULE expansion, and token-efficient output through a deterministic temporal computation layer (Truth Engine).

[Main capabilities]

• Temporal awareness: agents can call get_temporal_context to obtain the real time and time zone; resolve_datetime converts "next Tuesday at 2 PM" into an exact RFC 3339 timestamp, avoiding hallucinations.
• Atomic booking: first lock the time slot, verify there is no conflict, and then write. When two agents compete for the same time slot, exactly one succeeds and the other receives a clear error, avoiding double booking, using Two-Phase Commit.
• Computed availability: merge the free/busy data of multiple calendars into a unified view, letting AI see real availability rather than a raw event list that it must interpret itself.
• Deterministic RRULE expansion: handles daylight saving time transitions, BYSETPOS=-1 (the last weekday of each month), EXDATE with time zones, leap year recurrence, and edge cases such as INTERVAL>1 combined with BYDAY according to RFC 5545, without relying on LLM inference.
• Token-efficient output: TOON (Token-Oriented Object Notation) format compresses about 40% of the tokens compared with standard JSON, reducing cost and context window usage. TOON is theData tools have a default output format; JSON can be explicitly obtained via format: "json".
• Tool scale: 18 tools in total, divided into 5 layers. 15 core tools are always available, and 3 additional Open Scheduling tools are available only in platform mode. The 5 layers are: Layer 0 Discovery, Layer 1 Time Context, Layer 2 Calendar Operations, Layer 3 Availability, Layer 4 Booking.
• Protocol support: MCP, A2A, REST, browser.
• Framework integrations: LangGraph, CrewAI, OpenAI Agents SDK.

[How to Use] Two usage paths:

1. Individual users: after connecting a calendar, an AI agent handles availability checks, time zone resolution, and meeting booking. Compatible with Claude Desktop, Cursor, OpenClaw, Manus, and any MCP-compatible AI client. The installation command is npx @temporal-cortex/cortex-mcp, and the setup wizard interactively guides you through provider authentication, time zone configuration, and MCP client setup.
2. Developers/products: to add scheduling capabilities to AI agents or products, you can use the local MCP server (npx @temporal-cortex/cortex-mcp, full tool suite, zero infrastructure), the platform REST API (app.temporal-cortex.com, hosted, API key, usage dashboard, Open Scheduling network), or framework integrations. For the REST API reference, see temporal-cortex.com/docs/rest-api. Manual setup in three steps: install prerequisites (Node.js 18+ or Docker, and at least one calendar provider); add the MCP configuration to the AI client configuration file; run the authentication flow npx @temporal-cortex/cortex-mcp auth google (or outlook / caldav), which also configures the time zone, week start day, and telemetry preferences. Authentication: the simplest way is to run npx @temporal-cortex/cortex-mcp setup, which completes authentication, configuration, and MCP client setup in one go. When authenticating each provider separately, credentials are savedto ~/.config/temporal-cortex/credentials.json, and providers are registered in ~/.config/temporal-cortex/config.json. Multiple providers can be connected, and the server automatically discovers all configured providers at startup and merges calendars into a unified view. During authentication, interactive setup guidance is provided for: time zone (automatically detects the system time zone and provides a fuzzy search selector for all 597 IANA time zones), week start day (Monday, ISO standard, or Sunday), telemetry (optional anonymous data, off by default). Preferences are stored in ~/.config/temporal-cortex/config.json and can be overridden per session via the TIMEZONE and WEEK_START environment variables. Provider credential requirements: Google Calendar requires Google OAuth credentials; Microsoft Outlook requires an Azure AD app registration (MICROSOFT_CLIENT_ID); CalDAV (iCloud, Fastmail, etc.) requires an app-specific password. Platform mode: no local Node.js or local OAuth credentials required. Register at app.temporal-cortex.com, connect Google Calendar or Microsoft Outlook via OAuth through the dashboard, generate an API key, and then add the MCP configuration to the AI client. Platform capabilities include: no need to manage OAuth credentials, no need for Node.js (clients connect directly to the cloud endpoint via HTTP), usage dashboard, hosted calendar connections (token refresh, re-authentication, and provider health checks handled server-side), multi-agent coordination (distributed locks prevent concurrent booking conflicts), usage metering (track tool calls by agent and team), content firewall (automatic prompt injection detection and zero-width Unicode stripping), caller-based policies (enforce booking rules by agent, such as maximum duration, allowed time slots, booking limits). All 15 core tools and 5 layers behave consistently in both modes. Open Scheduling and Temporal Links: platform users can enable Open Scheduling, making availability publicly queryable by AI agents and humans without an API key. In the dashboard Set the slug in Settings > Scheduling and enable it, then share the Temporal Link (e.g., book.temporal-cortex.com/billy). Callers can access it through Agent Card discovery, REST availability queries, REST booking, A2A JSON-RPC calls to query_availability or book_slot, identity resolution, and other interfaces. The same link also works in a browser, where humans will see a booking page with date/time selectors and a form. Installation verification: Every GitHub Release publishes SHA256 checksums, which are embedded into the npm package as checksums.json for automatic verification after installation. The postinstall script downloads the platform-specific binary and compares the SHA256 hash; if it does not match, installation fails and shows the expected and actual hashes. The binaries are cross-platform compiled from Rust source via GitHub Actions, covering darwin-arm64, darwin-x64, linux-x64, linux-arm64, and win32-x64. The computation layer is open-sourced at temporal-cortex/core. It can be run via a Docker container for maximum isolation, without needing to install Node.js on the host machine.

[Use cases]

• Add autonomous scheduling capabilities to AI agents or products.
• Scenarios that require merging availability across multiple calendar providers and preventing double booking.
• Scheduling systems that need deterministic handling of complex recurrence rules (daylight saving time, last business day of the month, leap years, etc.).
• Individuals or teams who want to publicly expose availability and accept AI or human bookings through Open Scheduling.

[Limitations]

• Local mode requires Node.js 18+ (or Docker), as well as authentication credentials for at least one calendar provider.
• Different providers require separate OAuth credentials or app-specific passwords, and the specific steps are not fully expanded in the materials.
• The 3 additional Open Scheduling tools are available only in platform mode.
• Information such as pricing, rates, and quotas for platform mode is not yet provided in the materials.
• For specific tool input/output schemas and examples, see docs/tools.md; the materials do not list them item by item.