exo-explore/exo
40.7
Weak · 14 November 2026
73.2k
lines of production code
Python
with TypeScript, Swift, Rust
8
measurements over time
What this system is
Features
Add example configurations for local exo cluster integration
Added example configuration files for integrating with a local exo cluster. The new \claude\_code.sh\ script provides a shell environment to run Claude Code against a local endpoint using the GPT OSS 120B model, while \opencode.json\ supplies the corresponding JSON configuration for the OpenCode editor, defining the local API endpoint, model details, and context/output limits.
_tmp/config\examples · high confidence
Added PyInstaller build specification for Exo
The repository now includes a PyInstaller specification file (exo.spec) that defines how to package the Exo application into a standalone executable. This configuration ensures the inclusion of the dashboard build assets, resource files, shared model directories, and necessary MLX libraries, while also handling platform-specific requirements such as the macmon binary on macOS.
packaging/pyinstaller · high confidence
Custom macOS DMG installer with guided drag-to-Applications layout
The macOS packaging process now generates a polished DMG installer for EXO that features a custom background image and a specific window layout. The installer window displays the application icon on the left and an alias to the Applications folder on the right, connected by a curved arrow graphic to guide users in dragging the app to install. This change introduces the \create-dmg.sh\ build script and a Python helper (\generate-background.py\) to render the visual assets, replacing any previous default or unstyled DMG generation.
packaging/dmg · high confidence
EXO application entry point and core node initialization
The src/exo package now provides a standard Python entry point (\_\main\\_.py) that supports multiprocessing in frozen applications and initializes the core Node component. This includes setting up the Zenoh-based router, EventRouter, Master, Worker, and API services, handling signal-based shutdowns, and ensuring the models directory exists on startup.
src/exo · high confidence
EXO uninstaller script added with option to preserve models
A new uninstall script for EXO has been added, allowing users to cleanly remove system components such as the LaunchDaemon, network setup scripts, logs, and the EXO data directory. The script now supports a --keep-models flag, which preserves the \~/.exo/models directory during uninstallation, giving users control over whether their model data is retained or removed.
app/EXO · high confidence
Initial release of the EXO dashboard with chat, topology, and model management
This change introduces the core dashboard application for the EXO distributed AI cluster. It establishes the main layout with connection status banners and toast notifications, and implements the primary home page featuring a real-time cluster topology graph, a chat interface with model selection and message history, and a model picker modal. The UI includes support for mobile sidebars, download progress tracking, and specific hardware warnings for RDMA and macOS version mismatches. Also in this area: 1 commit (1 fix) fixing dashboard; 2 commits (0 fixes) modifying dashboard; 10 commits (9 fixes) fixing dashboard; Dashboard instance cards and model picker UX improvements; Improved failure diagnostics and onboarding clarity in the dashboard.
dashboard/src/routes · medium confidence · unverified
Initial repository scaffolding and developer guidelines
The repository has been initialized with foundational configuration and documentation files, including a Nix flake for the development environment, a justfile for build and lint commands, and a .python-version file specifying Python 3.13. To guide contributors and AI coding agents, the project includes AGENTS.md, CLAUDE.md, .cursorrules, and .clauderules, which enforce strict typing, pure functions, and Pydantic usage. Additionally, a README.md, PLATFORMS.md, and CONTRIBUTING.md have been added to document features, supported hardware (Apple Silicon, Linux), and contribution workflows.
(repo-wide) · high confidence
Initial support for FLUX and Qwen image generation models
The image generation engine now supports FLUX (including Schnell, Dev, and Kontext variants) and Qwen-Image models. This change introduces a new model adapter registry in the \src/exo/worker/engines/image/models\ package, allowing users to generate images using these specific model families. The implementation includes dedicated adapters for handling model-specific requirements, such as FLUX's dual text encoders and Qwen's attention masks, along with configuration for block types and synchronization steps.
src/exo/worker/engines/image/models · high confidence
Initial support for image generation via the M-Flux engine
The worker now includes a new image generation engine located in \src/exo/worker/engines/image\. This addition introduces the \DistributedImageModel\ class, which handles model loading, sharding, and diffusion execution, along with a \generate\_image\ function that processes generation tasks. Users can now generate images using the M-Flux backend, with support for configurable quality levels (low, medium, high), custom dimensions, seed control, and optional streaming of partial results.
src/exo/worker/engines/image · high confidence
Introduce Sparkle framework and commit Xcode project to version control
The EXO Xcode project is now tracked in the repository, enabling developers to build the app directly from source. This change adds the Sparkle framework as a dependency, which provides automatic update capabilities for the Mac app. Additionally, the commit scheme includes environment variables for AWS credentials, indicating integration with a bug-reporting or analytics service.
app/EXO/EXO.xcodeproj · high confidence
Introduce exo-bench for standardized inference throughput and energy measurement
The \bench\ area now includes a comprehensive benchmarking suite (\exo\_bench.py\) that measures inference throughput (prefill and generation tokens per second) and system-level resource consumption (GPU utilization, temperature, and power draw) for an exo cluster. The tool sends prompts to a dedicated \/bench/chat/completions\ endpoint, which disables the KV prefix cache by default and bans EOS tokens to ensure consistent, fair timing measurements. It supports concurrent request batching, configurable warmup runs, and prefix cache mode for specific optimization tracking. Results are output as JSON containing per-request statistics, cluster state, and time-series system metrics, with energy computed via trapezoidal integration of power samples. The suite is configured via TOML files (e.g., \bench.toml\, \single-m3-ultra.toml\) and includes specific benchmarks for disaggregated prefill-decode workloads (\prefill\_decode\_bench.py\) and tool-calling evaluation scenarios (\scenarios.toml\).
bench · high confidence
Introduce macOS menu bar application with cluster management and network diagnostics
The EXO macOS app is now available as a menu bar interface that manages the EXO process lifecycle, displays cluster topology and resource usage, and provides settings for custom namespaces, Hugging Face tokens, and environment variables. It includes a local network access checker that warns users when device discovery is blocked by macOS permissions and offers a direct link to System Settings, alongside a Thunderbolt Bridge service to prevent network loops. The app also supports automatic updates via Sparkle, launch-at-login registration, and a dedicated bug report window that uploads logs via presigned URLs. Also in this area: 3 commits (0 fixes) modifying app/EXO; Improved EXO process management and status visibility; New startup preferences and improved environment variable handling.
app/EXO/EXO · medium confidence · unverified
Introduce structured shared types for model backends, generation chunks, and node state
This change establishes a new \src/exo/shared/types\ module that defines the core data structures for the system. It adds a \Backend\ enum to identify supported inference engines (MlxMetal, MlxCpu, MlxCuda, Vllm) and a comprehensive \State\ model to track granular node information, including memory, disk, network, and Thunderbolt status. It also defines the \Chunk\ hierarchy for streaming responses, supporting token generation, tool calls, image generation, and prefill progress, alongside a \Command\ union for internal system operations like instance placement and downloads.
src/exo/shared/types · high confidence
Introduces MLX-based image generation engine with pipeline support
The worker now includes a new engine architecture for MLX-based image generation, featuring a base engine interface and a builder pattern to manage connection and loading phases. This update adds a dedicated image pipeline with a DiffusionRunner, block wrappers for joint transformer blocks, and a specialized ImagePatchKVCache to optimize key-value caching for image patches during generation. The worker's main loop and planning logic have been extended to handle ImageEdits and ImageGeneration tasks, including buffering input image chunks and managing an image cache for efficient processing.
src/exo/worker · high confidence
New API adapters for Claude, Ollama, and OpenAI Responses formats
The API layer now includes dedicated adapters for Claude Messages, Ollama Chat/Generate, and OpenAI Responses protocols, allowing users to interact with the model using these specific interface standards. These new modules handle the conversion of incoming requests from each format into the internal text generation parameters and translate internal model chunks back into the respective response formats, including support for multimodal inputs (images) and tool calls where applicable.
src/exo/api/adapters · high confidence
New API type definitions for OpenAI Responses, Claude, and Ollama compatibility
The \src/exo/api/types\ module has been introduced to centralize and standardize request/response schemas for external API compatibility. This change adds explicit Pydantic models for the OpenAI Responses API (including support for various tool calls like code interpreter, web search, and file search), the Claude Messages API (with support for thinking blocks and tool use), and the Ollama API (covering chat, generation, and model management endpoints). These types ensure that the API layer correctly handles wire-format data for these integrations, enabling users to interact with the system using familiar OpenAI, Claude, and Ollama client protocols.
src/exo/api/types · high confidence
New EXO macOS services for network configuration, state polling, and bug reporting
The app now includes a suite of new background services to improve stability and user experience. NetworkSetupHelper installs a LaunchDaemon to dynamically manage network locations and disable Thunderbolt Bridge to prevent packet storms, while ThunderboltBridgeDetector and ThunderboltBridgeService actively monitor for and prompt users to break detected bridge loops. ClusterStateService polls the local API for cluster state using an ephemeral, non-caching URLSession to prevent excessive disk I/O and memory usage. LocalNetworkChecker verifies actual local network connectivity via mDNS rather than relying solely on system settings. NetworkStatusService aggregates RDMA and interface status for the dashboard. Finally, BugReportService collects logs and state data and uploads them using presigned URLs.
app/EXO/EXO/Services · high confidence
New Integrations dashboard page with configuration helpers
A new Integrations page has been added to the dashboard, providing a centralized interface for configuring and connecting various AI models and tools. The page dynamically detects running models and offers pre-configured setup snippets for Anthropic Claude (including shell commands and JSON settings), OpenCode, and OpenAI Codex. It also includes state variables and default model assignments for Pi and OpenClaw integrations, streamlining the process of setting up these external services within the application.
dashboard/src/routes/integrations · high confidence
New LLM inference output parsing and chunking module
A new \llm\_inference\ package has been introduced to handle the parsing of model outputs and conversion into standardized response chunks. This module implements specific parsers for DeepSeek V3.2, DeepSeek V4, and GPT-OSS models, including logic to correctly count reasoning tokens for models that use thinking tags. It also handles tool call parsing and ensures that error finish reasons are properly propagated as error chunks to the user.
_src/exo/worker/runner/llm\inference · high confidence
New MLX generator engine with remote prefill and pipeline parallel support
This change introduces a new MLX-based text generation engine located in src/exo/worker/engines/mlx/generator. The engine implements core generation logic in generate.py, including support for pipeline parallel prefilling (overlapping stages across ranks) and remote prefilling via remote\_prefill.py (fetching KV cache chunks from a disaggregated server). It integrates with MLX's stream\_generate and logits processors, handles vision model embedding injection, and manages KV cache snapshots. This is a new component replacing or supplementing previous generation mechanisms.
src/exo/worker/engines/mlx/generator · high confidence
New Traces dashboard with deletion and Perfetto integration
The dashboard now includes a dedicated Traces page that lists all available traces, allowing users to select and delete multiple traces at once. Individual trace detail pages provide performance statistics broken down by phase and rank, along with options to download the raw trace data or open it directly in the Perfetto UI for advanced visualization.
dashboard/src/routes/traces · high confidence
New benchmarking, tracing, and internal type infrastructure
This change introduces new tooling and internal foundations for the exo platform. It adds two benchmarking scripts (\bench/exo\_eval.py\ and \bench/eval\_tool\_calls.py\) that allow users to evaluate model quality against standard datasets (GPQA, MMLU, AIME, HumanEval) and validate tool-calling behavior. It also adds a distributed tracing system (\src/exo/shared/tracing.py\) that exports Chrome-trace-compatible JSON for performance profiling, and a comprehensive set of internal type definitions (\src/exo/shared/types/\) for tasks, events, topology, and model cards to standardize the internal API.
exo · high confidence
New centralized state management and UI stores for the dashboard
The dashboard now uses a new set of Svelte stores to manage application state and UI interactions. The central app store handles topology data, chat state, and download progress, while dedicated stores manage user preferences like favorite and recently launched models (persisted to localStorage) and a global toast notification system for user feedback.
dashboard/src/lib/stores · high confidence
New integration test infrastructure and CLI tooling
This change introduces a new set of integration test scripts and a Python-based CLI tool (\exo\_tools\) to support testing and interacting with the exo cluster. The \tmp/old\_tests\ directory now contains shell scripts (\auto\_bench.sh\, \eval\_tool\_calls.sh\) and a Python helper (\get\_all\_models\_on\_cluster.py\) for automating benchmark runs, evaluating tool-call capabilities, and querying cluster state via Tailscale. Additionally, the \tools/src/exo\_tools\ package provides an HTTP client for the exo API (including streaming chat completions and state retrieval) and a harness for managing instance lifecycles, waiting for readiness, and handling model resolution.
_tmp/old\tests, tools · high confidence
New macOS native UI components for settings, bug reporting, and onboarding
The app introduces several new SwiftUI views and window controllers for macOS: a dedicated Bug Report window for submitting diagnostics, a First Launch Popout that guides new users to the dashboard, a native Settings window with tabs for General, Model, Advanced, Environment, and About sections, and new list/detail views for instances, nodes, and cluster topology. Also in this area: Improved EXO process management and status visibility; New startup preferences and improved environment variable handling.
app/EXO/EXO/Views · high confidence
New scripts for model card maintenance and cluster-wide model distribution
Added two new utility scripts to the scripts directory. The fetch\_kv\_heads.py script automates the process of fetching the num\_key\_value\_heads parameter from HuggingFace model configurations and updating the corresponding local TOML model card files, supporting both incremental updates for missing values and full overwrites. The download\_model\_to\_cluster.py script provides a way to bypass the standard placement logic to download a specified model to every node in an exo cluster simultaneously, handling model registration, topology discovery, and progress polling until all nodes report completion.
scripts · high confidence
New temporary utility scripts for model management, deployment, and security testing
Added a collection of temporary scripts in the tmp directory to support model operations and infrastructure. gen\_card.py automates the generation of inference model cards from Hugging Face. quantize\_and\_upload.py handles downloading, quantizing, and uploading models (including Qwen, FLUX, and FIBO variants) to Hugging Face. run\_exo\_on.sh facilitates deploying the EXO binary to remote hosts via SSH, using Zenoh for networking and checking availability on port 52415. run\_llm.py and run\_llm.sh provide client-side tools to stream chat completions from the local server. set\_rdma\_network\_config.sh configures macOS network settings for RDMA/Thunderbolt. Finally, test\_trust\_remote\_code\_attack.sh verifies that custom models are safely loaded with trust\_remote\_code disabled to prevent remote code execution.
tmp · high confidence
New topic-based routing system with network integration
The \src/exo/routing\ module has been introduced to provide a structured, topic-based messaging layer. This includes a \Router\ class that manages \TopicRouter\ instances, allowing components to subscribe to and publish messages on specific topics. The system integrates with the underlying networking layer (via \exo\_rs.NetworkingHandle\) to handle serialization, deserialization, and network transmission of messages, supporting policies like 'Always' or 'Minimal' publishing. It also handles connection state updates through \ConnectionMessage\ types. Also in this area: 1 commit (0 fixes) modifying exo, src/exo; 7 commits (0 fixes) modifying src/exo; 7 commits (1 fix) modifying src/exo; Improved cluster resilience and performance in event routing; New benchmarking, evaluation, and API type definitions.
src/exo/routing · medium confidence · unverified
New utility infrastructure for process management, observability, and data handling
This change introduces a comprehensive set of new utility modules in src/exo/utils that enhance the platform's runtime capabilities. It adds AsyncProcess for robust, asynchronous external process management with proper stdio capture and lifecycle handling, and ExitGuard to prevent shutdown hangs by forcing process exit after a grace period. For observability, it introduces PowerSampler to track and report time-weighted energy usage split between prefill and generation phases, and a startup banner that auto-opens the dashboard on first run. Data handling is improved with DiskEventLog for append-only, compressed event persistence, OrderedBuffer for event resequencing, and KeyedBackoff for per-key retry logic. Additionally, it provides new Pydantic base models (FrozenModel, TaggedModel) for stricter validation and serialization, along with supporting utilities for reactive variables, file system operations, and resource path resolution.
src/exo/utils · high confidence
Prefill/Decode disaggregation support for MLX models
Users can now enable prefill/decode disaggregation for MLX-based models by setting the ENABLE\_DISAGGREGATION cluster flag. This change introduces a new Advanced dashboard page (dashboard/src/routes/advanced) to manage this configuration, alongside the backend implementation in src/exo/worker/disaggregated and src/exo/worker/engines/mlx/disaggregated. The backend adds a TCP-based protocol and server to stream KV-cache chunks between nodes, an MLX-specific adapter to serialize/deserialize cache states, and a prefill service to handle remote prefill requests. Comprehensive end-to-end and unit tests are included to verify the protocol and adapter logic.
dashboard/src/routes/advanced, src/exo/worker/disaggregated · high confidence
Removals
Removal of initial networking and orchestration scaffolding
The initial scaffolding for distributed inference has been removed, specifically deleting the \networking\ and \orchestration\ modules. This eliminates the previous gRPC-based peer discovery, communication, and node management implementation, including the \GRPCDiscovery\, \GRPCServer\, \StandardNode\, and associated Protocol Buffer definitions. Users relying on this early distributed networking layer will no longer have access to these components.
networking, orchestration · high confidence
Removal of legacy MLX sharded inference components
The previous implementation of the MLX-based sharded inference engine has been removed. Specifically, the \InferenceEngine\ abstract base class, the \MLXFixedShardInferenceEngine\ concrete implementation, and the \Shard\ data class have been deleted from the codebase. This eliminates the fixed-shard processing logic and the associated shard definition structure that were previously used for local model inference.
inference · high confidence
Architecture
Core shared infrastructure and event-driven state management
The src/exo/shared package has been restructured to provide the foundational components for the application's state machine and configuration. This includes a new event-sourcing engine (apply.py) that processes domain events like task creation, deletion, and status updates to mutate the central State object, ensuring consistent state transitions. Configuration and file-system paths are now standardized via constants.py, adopting the XDG Base Directory Specification on Linux while respecting the EXO\HOME environment variable, and defining dedicated directories for logs, models, and caches. Additionally, the package introduces a new asynchronous election mechanism (election.py) for master node selection within the cluster, a robust logging setup (logging.py) with log rotation and compression, and an empty \\init\\_.py to formalize the package structure.
src/exo/shared · high confidence
Behavioural changes
14 commits (3 fixes) modifying .typings
A change to existing behaviour in .typings — 14 commits (3 fixs), 293 files: “fix(mlx): decoding a Qwen3.5 model no longer leaks until the runner crashes (\#2371)”; “Add Qwen3.8-Flash-Next (\#2513)”; “Nemotron-H: fix tensor and pipeline parallelism; add Nemotron 3.5 Lightning, Nemotron 3 Ultra and the Nemotron 3.5 Super-VL text path (\#2515)”; and 11 more (refs \#2371, \#2513, \#2515, \#2536, \#2511, …).
.typings · medium confidence · unverified
16 commits (4 fixes) modifying src/exo/worker/engines/mlx/models
A change to existing behaviour in src/exo/worker/engines/mlx/models — 16 commits (4 fixs), 162 files: “Add Qwen3.8-Flash-Next (\#2513)”; “Nemotron-H: fix tensor and pipeline parallelism; add Nemotron 3.5 Lightning, Nemotron 3 Ultra and the Nemotron 3.5 Super-VL text path (\#2515)”; “Add MiniMax M3 and M3.1-Flash (\#2536)”; and 13 more (refs \#2513, \#2515, \#2536, \#2577, \#2511, …).
src/exo/worker/engines/mlx/models · medium confidence · unverified
20 commits (1 fix) modifying app/EXO/EXOTests
A change to existing behaviour in app/EXO/EXOTests — 20 commits (1 fix), 15 files: “app: a welcome window for the first run (\#2495)”; “app: read download progress the way exo sends it (\#2552)”; “app: Uninstall EXO puts Thunderbolt Bridge back (\#2493)”; and 17 more (refs \#2495, \#2552, \#2493, \#2494, \#2590, …).
app/EXO/EXOTests · medium confidence · unverified
4 commits (0 fixes) modifying app/EXO/EXO/Views/Settings
A change to existing behaviour in app/EXO/EXO/Views/Settings — 4 commits, 8 files: “app: pass exo the app's build, and leave the cluster name to exo (\#2590)”; “app: move Settings to the standard Settings window, applied as changed (\#2496)”; “Connect only to nodes of the same build, and say so when one isn't (\#2588)”; and 1 more (refs \#2590, \#2496, \#2588, \#2497).
app/EXO/EXO/Views/Settings · medium confidence · unverified
4 commits (0 fixes) modifying rust
A change to existing behaviour in rust — 4 commits, 1 file: “rust: remove dead code”; “Rename exo\_pyo3\_bindings to exo\_rs (\#2131)”; “remove nightly (\#1538)”; and 1 more (refs \#2131, \#1538).
rust · medium confidence · unverified
6 commits (3 fixes) modifying src/exo/worker/engines/mlx/sharding
A change to existing behaviour in src/exo/worker/engines/mlx/sharding — 6 commits (3 fixs), 4 files: “Nemotron-H: fix tensor and pipeline parallelism; add Nemotron 3.5 Lightning, Nemotron 3 Ultra and the Nemotron 3.5 Super-VL text path (\#2515)”; “Add DeepSeek V4.1 Flash (\#2511)”; “deps: move mlx-lm to upstream ml-explore/mlx-lm (\#2509)”; and 3 more (refs \#2515, \#2511, \#2509, \#2512, \#2510, …).
src/exo/worker/engines/mlx/sharding · medium confidence · unverified
7 commits (0 fixes) modifying src/exo
A change to existing behaviour in src/exo — 7 commits, 79 files: “test: exo cluster behaviour on one machine (stub engine + multi-process harness) (\#2585)”; “A node that quits ends its models at once, but they move only if it doesn't come back (\#2603)”; “Keep the models users launch running: deployments and one reconciler (\#2601)”; and 4 more (refs \#2585, \#2603, \#2601, \#2602, \#2598, …). Also in this area: 1 commit (0 fixes) modifying exo, src/exo; 7 commits (1 fix) modifying src/exo; Added test coverage for API error handling, startup, and generation stream integrity.
_rust/exo\rs, src/exo/api/tests, src/exo/testkit · medium confidence · unverified
7 commits (1 fix) modifying src/exo
A change to existing behaviour in src/exo — 7 commits (1 fix), 48 files: “test(worker): wait for a cancellation to reach the runner's channel (\#2595)”; “Leave models that aren't on a node out of the cluster state (/state: 915 KB → 54 KB) (\#2592)”; “One membership module: Alive, Suspect or Gone per node, with every liveness timer in one place (\#2586)”; and 4 more (refs \#2595, \#2592, \#2586, \#2594, \#2593, …). Also in this area: 7 commits (0 fixes) modifying src/exo; Fix download path validation and reduce status announcement noise; Added tests for download cancellation, offline mode, and model directory management.
src/exo/download/tests · medium confidence · unverified
Dashboard UI overhaul with EXO Command Center theme and new chat capabilities
The dashboard has been visually redesigned with a new 'EXO Command Center' dark theme featuring yellow accents, CRT scanline effects, and a monospace font. Functionally, the chat interface now supports file attachments (images, text, PDFs) with drag-and-drop and paste handling, and includes a new model selector with smart auto-selection based on memory and task type. The chat experience also features uncertainty visualization via token-level heatmaps, a prefill progress bar with ETA, and the ability to cancel generation during the prefill phase.
dashboard/src/lib/components · high confidence
Enable MLX CPU builds and update Apple SDK to 26.2 on Nix
The Nix build environment now supports building MLX on CPU-only systems by enabling the MLX\_BUILD\_CPU option. Additionally, the Apple SDK overlay has been updated to version 26.2, and a custom Metal toolchain is provided to ensure compatibility with the newer SDK versions during compilation. Also in this area: Fix macOS build by replacing runtime detection with Nix-provided values.
nix · high confidence
Git LFS hooks added to enforce LFS usage
The repository now includes Git hooks (post-checkout, post-commit, post-merge, pre-push) that automatically invoke Git LFS commands. This ensures that large files tracked by Git LFS are properly managed during standard Git operations, and provides a clear error message if the 'git-lfs' tool is not installed on the user's system.
.githooks · high confidence
Improved context window handling for YARN-based models via RoPE patching
The MLX engine now applies a patch to the underlying \mlx\_lm\ library to align its YARN RoPE (Rotary Positional Embedding) implementation with vLLM's inverse-frequency blending formula. This change ensures better compatibility and stability for models using YARN scaling (such as Qwen3.5), specifically preventing infinite generation loops and memory leaks that could occur with the default implementation. The patch is applied automatically when the MLX engine initializes. Also in this area: MLX engine stability and performance improvements.
src/exo/worker/engines/mlx/patches · high confidence
Improved startup experience and exo process visibility
The app now provides clearer feedback during startup by displaying "Starting…" until the exo API responds, then switching to "Running" (ExoReadiness). It introduces a new menu bar icon style that visually indicates when exo is stopped, failed, or not responding via colored badges and greyscale indicators (MenuBarIconStyle). Additionally, the app now captures exo's stdout/stderr to a local log file (\~/.exo/exo\_app.log) with rotation support (ExoLogFile) and interprets unexpected exits to show user-friendly error messages in the UI (ExoExitReason). Also in this area: 3 commits (0 fixes) modifying app/EXO; 1 commit (0 fixes) modifying app/EXO.
app/EXO/EXO/Models · medium confidence · unverified
Introduce Nix-based build system and update API proxy port for the dashboard
The dashboard now uses a Nix build process (via dream2nix) to manage dependencies and generate the static site, which significantly improves build stability by decoupling the dependency resolution from source code changes. Additionally, the development server's API proxy configuration has been updated to forward requests to the backend service on port 52415, replacing the previous port 8000.
dashboard · high confidence
Introduce modular info gatherer with Thunderbolt bridge detection and network reachability checks
The \src/exo/utils/info\_gatherer\ module has been restructured into distinct components to improve system information gathering and network topology discovery. On macOS, the system now dynamically detects Thunderbolt Bridge network services by parsing \networksetup\ and \ifconfig\ outputs, replacing previous static or daemon-based approaches. Network reachability checks have been refactored to use \httpx\ with configurable timeouts and retry loops, verifying node identity via the API port. Additionally, a new \macmon\ integration provides detailed CPU, GPU, and memory metrics, while \system\_info\ handles OS version and hardware model detection. These changes enhance the reliability of node discovery and system profiling in distributed clusters.
_src/exo/utils/info\gatherer · high confidence
MLX engine refactored with distributed pipeline support and DeepSeek-V4 tool-calling fixes
The MLX engine has been restructured to support distributed pipeline execution and tensor sharding across multiple devices, introducing custom layer wrappers (PipelineFirstLayer, PipelineLastLayer) to handle inter-rank communication and cache synchronization during prefill and decode phases. This change also adds robust support for DeepSeek-V4 and V3.2 models, including specific DSML encoding/decoding logic to correctly handle tool calls and interleaved reasoning blocks, while fixing race conditions in distributed initialization and memory leaks in the KV prefix cache.
src/exo/worker/engines/mlx · high confidence
Networking stack migrated from libp2p to Zenoh
The Rust networking module has been replaced with a new implementation based on the Zenoh protocol. This change introduces a custom IPv6 multicast-based discovery mechanism (in \discovery.rs\) to locate peers and a compatibility layer (in \swarm.rs\) that maps the previous pub/sub and subscription operations to Zenoh's publish/subscribe and liveliness APIs. Users will now connect to peers via TCP using Zenoh's internal routing, replacing the previous libp2p-based transport and discovery logic. Also in this area: 3 commits (0 fixes) modifying rust/networking; Improved logging for incompatible network peers; Improved network resilience and discovery reliability.
rust/networking · medium confidence · unverified
New download coordinator with pause, cancel, and offline support
The download system has been restructured around a new DownloadCoordinator that manages model downloads at the node level. Users can now pause and delete active downloads, and downloads are automatically cancelled when the coordinator shuts down. An explicit offline mode is supported, allowing downloads to be skipped when the internet is unavailable, and the system now validates model IDs to prevent invalid folder names. Additionally, the coordinator handles read-only model directories for pre-downloaded models and improves progress reporting by throttling events and handling broken resource errors gracefully. Also in this area: 3 commits (1 fix) modifying src/exo/download.
src/exo/download · medium confidence · unverified
New master node orchestrator with disk-based event logging and intelligent instance placement
The master node component has been restructured into a new modular architecture (src/exo/master) that introduces a central orchestrator for managing model instances and tasks. This change implements disk-based event logging to replace unbounded in-memory storage, ensuring durability and preventing memory exhaustion during long-running sessions. The master now features an intelligent placement engine that assigns model shards to nodes based on available memory, network topology (including RDMA and Ethernet prioritization), and download progress, supporting both tensor and pipeline parallelism for models like DeepSeek and Gemma 4. Additionally, the system now supports image generation tasks and seamless chat UX with auto model selection, while improving load balancing by routing requests based on in-flight tasks rather than completed ones.
src/exo/master · high confidence
New typed response and shard metadata models for worker communication
This change introduces a new set of Pydantic-based type definitions in the worker shared types module to standardize data exchange between the worker and the API. For responses, it defines specific classes for text generation (including token-level logprobs and thinking flags), image generation (supporting partial streaming and multiple images), tool calls, and prefill progress, replacing previous ad-hoc structures. For model execution, it replaces the old \Shard\ object with a new \BaseShardMetadata\ hierarchy that explicitly distinguishes between Tensor, Pipeline, and CFG (Classifier-Free Guidance) parallelism shards, providing structured metadata for device placement and layer ranges.
src/exo/shared/types/worker · high confidence
Nix-based Python environment with CUDA and platform-specific build support
The Python build environment is now managed via a new \python/parts.nix\ file using \uv2nix\ and \pyproject-nix\. This change introduces support for Python 3.13 and integrates CUDA libraries (including CUDA 13) for Linux builds, while applying specific build fixes and dependencies for macOS (Darwin) using Apple SDK 26. It also ensures type stubs for the \exo-rs\ package are included to support static type checking with basedpyright.
python · high confidence
PDF support and Safari compatibility fix in dashboard file handling
The dashboard now supports uploading and processing PDF files alongside images, text, and audio. To ensure this works on Safari, a polyfill for ReadableStream's async iterator is included to fix PDF text extraction failures, and the pdfjs-dist worker is configured via CDN.
dashboard/src/lib/types · high confidence
Redesigned downloads page with model×node table view
The downloads interface has been completely redesigned to present download status in a model-by-node table layout. This new view allows users to see the state of downloads (completed, downloading, pending, failed) across multiple nodes simultaneously, displaying detailed progress metrics such as percentage, speed, and estimated time of arrival. It also integrates disk usage information for each node and provides controls to pause, resume, or delete active downloads directly from the UI.
dashboard/src/routes/downloads · high confidence
Structured runner diagnostics and robust stdio logging
The worker runner now captures and classifies stderr output into structured diagnostics (such as Metal GPU timeouts and ring transport errors) and redirects both stdout and stderr to persistent log files. This change improves observability and error recovery by allowing the supervisor to correlate specific failure patterns with runner status, rather than relying solely on unstructured logs.
src/exo/worker/runner · high confidence
Fixes
10 commits (9 fixes) fixing dashboard
A fix in dashboard — 10 commits (9 fixs), 28 files: “fix(dashboard): keep "model stopped" notices until dismissed, and tell a restart of this device apart (\#2570)”; “test(dashboard): run unit tests with node --test (\#2564)”; “fix(dashboard): only warn "RDMA NOT ENABLED" for devices cabled with Thunderbolt 5 (\#2573)”; and 7 more (refs \#2570, \#2564, \#2573, \#2572, \#2569, …). Also in this area: 1 commit (1 fix) fixing dashboard; 2 commits (0 fixes) modifying dashboard; Improved error handling, clipboard support, and download status utilities in the dashboard.
dashboard/src/lib/utils, dashboard/tests · medium confidence · unverified
API reliability, resource management, and startup fixes
The API now handles non-streaming errors correctly by returning a 500 status with an error body instead of an empty 200, and ensures interrupted or failed requests are properly terminated with errors rather than hanging or truncating. To prevent resource leaks, the API now resends 'TaskFinished' commands if the master changes, keeps running models through master changes, and correctly reports task completion when clients disconnect. Startup is more robust: the API binds the port early to fail cleanly if it is already in use, and the ready banner is printed only after the API is actually serving. Additionally, the in-memory event log for /events is now bounded by count and image data size to prevent unbounded disk growth, and SSE connections are kept alive to prevent timeouts during long prefill phases.
src/exo/api · high confidence
Test coverage
Added comprehensive test suite for MLX worker engine; Added test coverage for shared utilities; Added test infrastructure and test plan for the worker module; Added tests for MLX batch generation equivalence and logprob extraction; Added tests for async process, channels, event log, PID file, power sampling, and tagged unions; Added tests for routing components; Added tests for state application logic; Added unit tests for image generation cancellation logic; Added unit tests for runner output parsing, event ordering, and supervisor error handling; Added unit tests for the worker plan module; Added unit tests for worker lifecycle and state management; Initial test suite for master node logic, placement, and topology; Integration test infrastructure and test suite.
Dependencies
Initial dependency manifests for Rust, Python, and Dashboard workspaces
This change introduces the foundational dependency configuration files for the project's multi-language structure. It adds a Rust workspace (Cargo.toml/Cargo.lock) that establishes the \exo\_rs\ and \networking\ crates, utilizing Zenoh for networking and PyO3 for Python bindings. It also adds Python project manifests (\pyproject.toml\) for the main application, benchmarking tool, and shared tools, defining dependencies such as FastAPI, MLX, and Transformers. Additionally, it includes the Svelte-based dashboard configuration (\package.json\/\package-lock.json\) and the macOS app's Swift package resolution file.
(dependencies) · high confidence
Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
How this codebase got here
Score
- CAI 38 → 41 (+2.5)
- Rubric changed (rubric-2026.10.5 → rubric-2026.11.4) — scores are not directly comparable.
Lenses
- Code Health 28 → 29 (+1.1)
- Architecture 93 → 92 (-0.5)
- Maturity 64 → 68 (+4.7)
- Readiness 33 → 37 (+4.2)
- Security 51 → 51 (+0.1)
- Domain Modelling 100 (new)
- Accessibility 51 → 56 (+4.8)
- Performance 100 → 100 (+0.0)
Resolved (104)
- API._apply_state (cognitive 26) (src/exo/api/main.py)
- API.get_placement_previews (cyclomatic 17) (src/exo/api/main.py)
- Change coupling: chat_completions.py ↔ utils_mlx.py (src/exo/api/adapters/chat_completions.py)
- Change coupling: responses.py ↔ utils_mlx.py (src/exo/api/adapters/responses.py)
- ClassTooLong: ContentView (app/EXO/EXO/ContentView.swift)
- ClassTooLong: SettingsView (app/EXO/EXO/Views/SettingsView.swift)
- ContentView.clusterThunderboltBridgeView (cognitive 19) (app/EXO/EXO/ContentView.swift)
- DownloadCoordinator._emit_existing_download_progress (cognitive 45) (src/exo/download/coordinator.py)
- DownloadCoordinator._emit_existing_download_progress (cyclomatic 18) (src/exo/download/coordinator.py)
- Duplicate initialization logic for discovery. Discovery.new takes explicit ports and a ZenohId, while networking.open takes a Config and a namespace string. It is unclear if networking.open internally instantiates Discovery or if the user is expected to instantiate Discovery manually. The parameter types (ZenohId vs str for identity, [u8; 8] namespace vs str namespace) are inconsistent across the API surface.
- Duplicated block (10 lines × 2) (app/EXO/EXO/ContentView.swift)
- Duplicated block (10 lines × 2) (src/exo/worker/engines/mlx/auto_parallel.py)
- Duplicated block (10–12 lines × 2) (src/exo/download/download_utils.py)
- Duplicated block (10–13 lines × 3) (app/EXO/EXO/Views/SettingsView.swift)
- Duplicated block (12 lines × 2) (app/EXO/EXO/ContentView.swift)
- Duplicated block (13 lines × 2) (app/EXO/EXO/Services/ClusterStateService.swift)
- Duplicated block (13 lines × 2) (src/exo/api/main.py)
- Duplicated block (14 lines × 2) (app/EXO/EXO/ContentView.swift)
- Duplicated block (14 lines × 2) (src/exo/worker/engines/mlx/auto_parallel.py)
- Duplicated block (15–16 lines × 4) (app/EXO/EXO/ContentView.swift)
- …and 84 more
New (215)
- Change coupling: EXOApp.swift ↔ NetworkSetupHelper.swift (app/EXO/EXO/EXOApp.swift)
- Change coupling: ExoProcessController.swift ↔ SettingsView.swift (app/EXO/EXO/ExoProcessController.swift)
- Change coupling: claude.py ↔ generate.py (src/exo/api/adapters/claude.py)
- Change coupling: commands.py ↔ events.py (src/exo/shared/types/commands.py)
- Change-coupling hub: utils_mlx.py → chat_completions.py, claude.py, responses.py (src/exo/worker/engines/mlx/utils_mlx.py)
- Cluster._stream_chat (cognitive 22) (src/exo/testkit/cluster.py)
- Cluster.close (cognitive 19) (src/exo/testkit/cluster.py)
- Cluster.violations (cognitive 38) (src/exo/testkit/cluster.py)
- Cluster.violations (cyclomatic 17) (src/exo/testkit/cluster.py)
- Compressor.call (cognitive 16) (src/exo/worker/engines/mlx/models/deepseek_v4/model.py)
- Coverage not measured — JavaScript/TypeScript suite
- DeepseekV4Cache.accumulate_windows (cognitive 67) (src/exo/worker/engines/mlx/models/deepseek_v4/model.py)
- DeepseekV4Cache.accumulate_windows (cyclomatic 36) (src/exo/worker/engines/mlx/models/deepseek_v4/model.py)
- DeepseekV4Cache.prepare (cognitive 17) (src/exo/worker/engines/mlx/models/deepseek_v4/model.py)
- DeepseekV4Cache.update_pool (cognitive 16) (src/exo/worker/engines/mlx/models/deepseek_v4/model.py)
- DiagnosticsSummary.make (cognitive 27) (app/EXO/EXO/Views/Settings/DiagnosticsView.swift)
- DiagnosticsSummary.make (cyclomatic 16) (app/EXO/EXO/Views/Settings/DiagnosticsView.swift)
- Duplicated block (10 lines × 2) (src/exo/worker/engines/mlx/models/deepseek_v4/encoding.py)
- Duplicated block (10 lines × 2) (src/exo/worker/engines/mlx/models/glm/glm5_next.py)
- Duplicated block (10 lines × 2) (src/exo/worker/engines/mlx/models/qwen3/sharding.py)
- …and 195 more
Changes since last survey
- 81 commits — 50 feature/other, 31 fixes
By area
- src/exo — 52 commits
- dashboard/src — 14 commits
- app/EXO — 10 commits
- (root) — 1 commit
- dashboard/package.json — 1 commit
- docs/imgs — 1 commit
- resources/inference_model_cards — 1 commit
- rust/networking — 1 commit
Notable commits
- fix: Add GLM-5.3, and fix GLM-5.2/5.3 output past 2048 tokens (#2512)
- fix: Fix DeepSeek V4 against the official reference and add the 0731/0813 release builds (#2516)
- fix: Nemotron-H: fix tensor and pipeline parallelism; add Nemotron 3.5 Lightning, Nemotron 3 Ultra and the Nemotron 3.5 Super-VL text path (#2515)
- fix: app: say what a bug report sends, and send only what triage needs (#2498)
- fix: fix(api): send launches and other cluster changes again until they show, also to a new master (#2581)
- fix: fix(dashboard): count every device in a model's download total (#2567)
- fix: fix(dashboard): don't call a model Ready when its devices can't serve it (#2566)
- fix: fix(dashboard): don't lose conversations when two tabs are open (#2569)
- fix: fix(dashboard): don't recommend models the devices have no disk space for (#2571)
- fix: fix(dashboard): keep "model stopped" notices until dismissed, and tell a restart of this device apart (#2570)
- fix: fix(dashboard): keep the partial reply when a stream fails (#2565)
- fix: fix(dashboard): only warn "RDMA NOT ENABLED" for devices cabled with Thunderbolt 5 (#2573)
- fix: fix(dashboard): say a model is running before deleting its files (#2572)
- fix: fix(dashboard): show stalled downloads, and no ETA until it's known (#2568)
- fix: fix(dashboard): show what exo does with a launched model, and keep what the user chose (#2604)
- fix: fix(master): apply its own events as it decides on them; a launch that doesn't fit is refused (#2584)
- fix: fix(mlx): wait for a collective on the CPU before the GPU uses it (#2510)
- fix: fix(placement): check each node's share against the memory a model can use there (#2486)
- fix: fix(utils): close each end of an mp channel once, even from two threads (#2554)
- fix: fix(worker): a runner stuck on its way to serving is stopped, not waited on forever (#2501)
- …and 61 more
Architecture
- Unchanged — 0 containers · 1 contexts · 0 edges
Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.
Survey your own repository
exo-explore/exo was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.
About this page
- The score is its most recent published measurement, taken on 14 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
- Measured at commit a60b95698ca81737d367443674dd5563cbc9559b — the exact code this score is about.
- Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
- Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.65+preprod-36d4cd83012f.