​
v0.3.4 (Current)

Release on 6 October 2026.

​
Added

  • Install script — curl -fsSL https://loop.soket.ai/install | bash installs the latest release to ~/.loop/bin, verifies .sha256, skips a version that is already present, and can pin --version 0.3.4 or skip PATH edits
  • Skill toggles — /skills picker, /skills enable and /skills disable, saved as disabledSkills. Discovery includes ~/.agents/skills, ~/.claude/skills, and trusted project skill directories
  • Skill activation — /skill:<name> keeps a skill in the system-prompt catalog until /new, including skills with disable-model-invocation
  • MCP toggles and OAuth — /mcp picker, enable and disable (disabledMcpServers), and /mcp login / /mcp logout for remote HTTP servers (authorization code with PKCE)
  • --mcp-host — Interface for --serve-mcp (default 127.0.0.1). A non-loopback address requires --mcp-token

​
Changed

  • Soket provider — API base URL is https://api.soket.ai/v1. /login soket points at free credits on console.soket.ai. Key lookup order is SOKET_API_KEY, then TENSORSTUDIO_API_KEY, then LOOP_API_KEY
  • Linux releases — Built on Ubuntu 22.04 (glibc). musl distributions such as Alpine are not supported by the published archives
  • Context files — Each directory can contribute AGENTS.md, CLAUDE.md, CLAUDE.local.md, and .claude/CLAUDE.md. The agent directory and ~/.claude/CLAUDE.md are included

​
Fixed

  • Markdown tables in the streaming TUI
  • Multiline paste no longer underflows on character offsets
  • Up-arrow history moves through slash commands
  • Installer quote bug, and a missing checksum now stops the install unless LOOP_ALLOW_UNVERIFIED=1

​
v0.3.1

​
Added

  • MCP in the default binary — /mcp and --serve-mcp ship in the release build with no extra Cargo features
  • Optional telemetry feature — The default binary no longer links OpenTelemetry. /tracing and --trace-* need --features telemetry
  • /tracing — Langfuse or any OTLP/HTTP collector, with a setup wizard. Secrets stay in the credential store
  • /login presets — Soket, OpenRouter, OpenAI, and a custom OpenAI-compatible endpoint. First launch opens the picker when no key is set
  • /ttfb — Response-header timeout (responseHeaderTimeoutMs, default 60s, 0 waits indefinitely)
  • /workflow CLI integration — Plan a multi-agent task graph from a goal and stream live WorkflowTask cards in the TUI (--features orchestration)
  • Workflow progress channel — WorkflowProgressEvent + harness progress_tx for task started / completed / failed UI updates
  • Richer WorkflowResult — failed_tasks (id + error) and total_task_count alongside completed task_results
  • WorkflowEngine::subscribe() — Live WorkflowEvent broadcast without polling the event log
  • Desktop app (loop-desktop) — GPUI sessions sidebar, streaming chat, diff review, PTY dock; shared harness via loop-app-core
  • loop-app-core crate — Shared bootstrap, config, theme, and tool-approval wiring for CLI + desktop

​
Improved

  • Auto-filled PlannerContext — When omitted, harness supplies cwd + available tool names
  • Soft failure detection — AgentWorker fails tasks that narrate blockage even when the agent loop returned Ok
  • Empty tool-filter fallback — Keep the full toolset if a planner filter matches nothing
  • File artifact collection — Successful write results verified on disk become workflow Artifacts
  • Ctrl+O — Expand/collapse tool, thinking, and workflow task detail

​
Architecture

  • Cargo workspace: loop-ai, loop-agent, loop-app-core, loop-cli, loop-desktop, loop-mcp, plus optional loop-orchestration and loop-telemetry
  • Dependency flow: surfaces → loop-app-core → loop-agent → loop-ai; orchestration and telemetry are feature-gated

​
v0.2.0

​
Added

  • Multi-agent orchestration — Temporal-inspired workflow engine with planner, scheduler, and shared memory
  • MCP server mode — --serve-mcp exposes Loop tools over HTTP via Model Context Protocol
  • MCP client — Connect external MCP servers and bridge their tools into agent sessions
  • Container sandbox — Podman-based isolation with full/partial modes and multiple OCI runtimes (runc, crun, gVisor, krun)
  • Session trees — Fork, branch, clone, and navigate conversation histories
  • SQLite sessions — Full-text search, durable storage, and concurrent access
  • Skills system — Reusable agent capabilities via SKILL.md files
  • Prompt templates — Slash-command templates with argument substitution
  • Rhai extensions — Custom commands and notifications via Rhai scripting
  • Themes — Customizable TUI themes with light/dark mode
  • Declarative hooks — JSON lifecycle hooks for agent events
  • Context files — Auto-load AGENTS.md / CLAUDE.md from project directories
  • Thinking modes — Configurable chain-of-thought reasoning levels
  • Context compaction — Automatic context-window management
  • Token tracking — Per-session cost estimation and cache-waste metrics
  • Headless mode — --print flag for scripting and CI pipelines
  • Project trust — Security gate for project-level configuration overrides
  • Command safety — Blocklist for dangerous shell commands
  • File edit review — Configurable diff review in Cursor/VS Code

​
Architecture

  • Cargo workspace with focused crates
  • Unified LLM API layer (loop-ai) with provider abstraction
  • Event-sourced orchestration engine (loop-orchestration)
  • Trait-based extension points (SessionStore, SandboxFactory, ToolProvider, Worker)
  • Multi-platform release builds (Linux x86_64/ARM64, macOS Intel/Apple Silicon, Windows x64)
View the latest changes in the TUI with the /changelog slash command.