Tyrell

Autonomous fleet orchestration, agent telemetry, multi-provider quota governance, and cross-harness coordination for macOS.

1. Executive Pitch

As multi-agent artificial intelligence development accelerates, software teams deploy dozens of autonomous coding agents concurrently across workstation clusters. Without central visibility, these environments rapidly encounter fragmented session histories, uncoordinated API spend that triggers aggressive provider rate limits, silent background daemon crashes, and dropped agent communications.

Tyrell is a native macOS orchestration platform and fleet observatory designed from the ground up to solve multi-agent coordination. Built in modern Swift and SwiftUI with a Universal2 (Apple Silicon and Intel) engine, Tyrell connects distributed machines into a cohesive, observable execution environment. It surfaces real-time host telemetry, aggregates cross-agent conversation buses, monitors multi-provider rolling rate limits with second-level accuracy, provides advisory token-lease governance, and indexes sessions across every major agent runtime—all while upholding zero-trust authentication and local privacy invariants.

2. Product Overview

Tyrell operates across four tightly integrated planes:

1. The Native Management Console (Tyrell.app & TyrellBar.app): A Liquid Glass and Signal Glass desktop interface providing real-time fleet health verdicts, interactive 24-hour activity heatmaps, threaded agent messaging, provider usage gauges, and node inventories.
2. The Autonomous Service Daemon (tyrelld): A resilient, self-healing background daemon managing peer discovery, SQLite write-ahead logging (WAL), durable message queues, and cross-host synchronization.
3. The Developer & Agent Interfaces (tyrell CLI & tyrell-mcp): Fast command-line utilities and a standardized Model Context Protocol (MCP) server enabling both human operators and autonomous agents to query fleet health, search indexes, trigger scans, and check budget advice.
4. The Resilient Coordination Plane: Decentralized hub election, automatic Bonjour and secure mesh discovery, leased failover, and hub-to-client service configuration management.

3. Key Features

Unified Agent Chat & Fleet Coordination Bus

• Multi-Agent Message Hub: Direct visibility into the asynchronous maildir agent bus, grouping thousands of automated messages into coherent, topical conversations.
• Cross-Harness Search ("My Threads"): Centralized indexing and search across all conversations initiated by engineers across Claude Code, Codex, Antigravity, Grok, and GitHub Copilot.
• Actionable Thread Management: Instant filtering by participant, host, priority, time window, and ticket linkage, complete with in-app message response composition.
• Privacy & Redaction Boundary: Automatic redaction of secrets, keys, and private credentials before indexing or display.

Multi-Host Fleet Heatmap & Signal Glass Observatory

• 24-Hour Activity Heatmap: Granular visualization of file index updates, agent activity, and task executions across all cluster nodes.
• Real-Time Node Freshness: Continuous tracking of node heartbeat manifests, total indexed files, managed storage size, and mirror sync queues.
• Verdict-First Status: Instant high-level system assessments ("Nothing needs you" vs. actionable alerts) preventing operational alert fatigue.

Live Multi-Provider Quota Governance & Advisory Slots

• Calibrated Rolling Windows: Continuous tracking of provider usage limits across Anthropic, OpenAI, Google Gemini, OpenRouter, and GitHub Copilot (5-hour burst, 7-day rolling, and monthly quotas).
• Proactive Anti-Stall Advice: Automatic grading of usage buckets (proceed, conserve, hold) providing definitive guidance to autonomous agent dispatchers before rate limits trigger.
• Advisory Slot Leasing Engine: Quota-aware reservation locks preventing agent swarms from exhausting collective API tier allocations during parallel fan-out.
• Standing Budget Holds: Dedicated mechanisms allowing operators to pause API consumption fleet-wide without terminating existing workloads.

Autonomous Hub Election & Service Inheritance

• Decentralized Leader Election: Dynamic election of the master coordination node based on hardware class (desktop vs. laptop), AC power status, network stability, and historical availability.
• Leased Failover: Secure 600-second leases with automatic renewal every 180 seconds, ensuring graceful recovery if the primary hub goes offline.
• Service Inheritance: Automatic advertisement and propagation of shared development services across connected clients, updating agent tool configurations cleanly.

Native Background Service Management

• Zero-Drift LaunchAgent Lifecycle: Native in-app and CLI controls (tyrell service register|unregister|restart) enforcing a single stable binary path to protect Full Disk Access and macOS TCC approvals.
• Flexible Execution Modes: Choose between continuous background execution (login item) or execution strictly bound to the desktop application lifetime.

4. Architectural Boundaries

Tyrell maintains strict separation of concerns within the developer ecosystem:

┌─────────────────────────────────────────────────────────────┐
│                        TYRELL                               │
│  - Fleet Telemetry & Observatory                            │
│  - Agent Coordination & Chat Bus                            │
│  - Multi-Provider Quota Intelligence & Advisory Slots        │
│  - Hub Election & Service Inheritance                       │
└──────────────────────────────┬──────────────────────────────┘
                               │
               Advisory Quota Telemetry (Read-Only)
                               ▼
┌─────────────────────────────────────────────────────────────┐
│                    RDMODELROUTER                            │
│  (Independent Product — Model Routing & Dispatch)           │
│  - Account & Provider Switching                             │
│  - Token Bucket Routing                                     │
│  - Model Fallback & Multiplexing                            │
└─────────────────────────────────────────────────────────────┘

[!IMPORTANT] Tyrell is not a model router or LiteLLM gateway. Model selection, provider switching, and request dispatch are strictly managed by rdmodelrouter. Tyrell supplies authoritative, read-only usage evidence and advisory slot reservations that routing tools consume.

5. Platforms & System Requirements

6. Status & Availability

• Current Version: 0.2.0 (Build 46)
• Git Milestone: Commit 85b5dda
• Release Status: Active Fleet Production
• Operational Priority: Standing Priority (active development and iteration; overrode legacy maintain tier on 2026-09-26)
• Availability: Private enterprise / infrastructure deployment across authenticated cluster hosts. Not available on public app stores.

7. Pricing

• Model: Internal infrastructure platform.
• Retail Pricing: None / Undecided (planned for external release).

8. Documentation & Technical Links

• Tyrell Product Requirements Document (Internal)
• Tyrell Interface Specification (CLI, MCP, API) (internal; see PRD)
• Service Mode Decision Record (internal; see PRD)
• Hub Election & Failover Architecture (internal; see PRD)
• Unified Chat Inbox PRD (internal; see PRD)
• Live Screen Captures & Captions
• Brand & Web Assets Specification