Introduction

SwarmClaw is a self-hosted AI agent runtime and control plane. It gives you one place to run agents, connect multiple model providers, delegate work across agents and CLI backends, automate recurring work, manage long-term context, and bridge your agents into external systems.

Features

  • 24+ built-in providers: Claude Code CLI, Codex CLI, OpenCode CLI, Gemini CLI, GitHub Copilot CLI, Cursor Agent CLI, Qwen Code CLI, Anthropic, OpenAI, OpenRouter, Google Gemini, DeepSeek, Groq, Together, Mistral, xAI, Fireworks, Nebius, DeepInfra, Ollama, LM Studio, OpenClaw, Goose, and Hermes Agent, plus compatible custom endpoints.
  • Agent builder: personalities, system prompts, tool selection, runtime skills, OpenClaw routing, gateway preferences, and MCP integrations.
  • Delegation: agent-to-agent delegation plus CLI delegation to Claude Code, Codex, OpenCode, Gemini CLI, Cursor Agent CLI, Qwen Code CLI, and other configured backends.
  • Projects and goals: durable project scope, goal hierarchy, linked agents/tasks/schedules, and shared operating context.
  • Tasks and schedules: queue work, model dependencies, archive and restore schedules, and review run history from operational surfaces.
  • Structured Sessions: bounded multi-agent runs with templates, durable event history, artifacts, and optional live rooms.
  • Connectors and Inbox: connect chat platforms and local file queues, keep external sender threads isolated, and review connector conversations outside the main chat surface.
  • Memory and knowledge: long-term memory, shared knowledge sources, knowledge grounding, citations, hygiene, and source lifecycle management.
  • MCP servers and extensions: inject external MCP tools into agents and install external extensions for hooks, UI, providers, or connectors.
  • Autonomy and activity: estops, incidents, reflections, audits, usage rollups, and background daemon controls.
  • Hosted deploy + OTLP foundation: root-level Render, Fly.io, and Railway deploy configs plus opt-in OTLP trace export for operators.

Core Model

SurfaceBest For
Direct chatopen-ended work with one agent
Chatroompersistent free-form multi-agent collaboration
Structured Sessionbounded runs with outputs, operator controls, and completion
Taskconcrete queued work with ownership, dependencies, and retry history
Projectdurable scope that groups agents, tasks, schedules, and goals
Inboxisolated external sender conversations from connectors
Memorydurable recall tied to agents, chats, and learned behavior
Knowledgeshared sources, ingestion, grounding, and citations across the workspace

What Changed Recently

  • v1.9.35 fixes fresh npm-global fallback builds by shipping the OpenClaw media proxy's mime-types runtime and type declarations as package dependencies.
  • v1.9.34 recovers credential secrets from prior npm-global build env files, surfaces connector credential decrypt failures clearly, and makes enabled external extension tools attachable to scoped agents.
  • v1.9.33 fixes execute credential injection, persists credential secrets across npm-global upgrades, auto-queues agent-delegated tasks, adds failed-task retry, strips internal metadata before connector replies, preserves agent filesystem settings, and updates OpenClaw gateway routing to protocol v4.
  • v1.9.32 adds per-agent dream-model overrides, auto-compaction model routing, reflection memory quality gates, optional embedding dedup for reflection notes, and current ClawHub install guidance.
  • v1.9.27 fixes Intel macOS desktop native-module packaging, restores OpenRouter/provider saves that include UI metadata fields, and makes the downloads page avoid guessing Apple Silicon when Mac architecture is hidden.
  • v1.9.26 keeps empty successful LLM turns silent instead of rewriting them as user-visible errors, while preserving explicit provider and stream error reporting.
  • v1.9.25 adds gateway lifecycle controls for saved OpenClaw profiles, exposes CLI lifecycle actions, skips draining or cordoned gateways during route selection, surfaces lifecycle attention in Operations Pulse, and fixes Slack peer-bot collaboration in shared channels.
  • v1.9.23 repairs stale cron timing before schedules skip the nearest run, advances recurring schedules from scheduler tick time, adds deterministic stagger offsets, and keeps schedule-created board tasks attached to a persistent mission link.
  • v1.9.22 adds direct web_extract and web_crawl tools with source-grounded page extraction, bounded same-origin crawls, routing support, and regression coverage.
  • v1.9.21 adds provider connection diagnostics with endpoint, model discovery, chat/gateway check timelines, and macOS signing workflow support for notarized desktop releases when credentials are configured.
  • v1.9.20 adds first-class LM Studio support, normalizes local OpenAI-compatible endpoints to /v1, and keeps stale local endpoints from overriding fixed cloud providers.
  • v1.9.19 strips repeated and malformed internal metadata before final assistant responses are stored, streamed, or sent through connectors.
  • v1.9.18 adds schedule preflight with server-backed next-run forecasts, explicit cron timezones, stagger controls, run-once delay controls, and CLI preview access.
  • v1.9.17 adds agent configuration history in the agent editor with recent saved versions, one-click restore, and regression coverage for list/restore behavior.
  • v1.9.16 adds agent planning controls, runtime prompt coverage, and agent-pack portability for strict planning mode.
  • v1.9.15 adds run handoff packets with structured JSON, copyable markdown, Run Review copy access, CLI access, and readiness guidance for completed, failed, queued, or running executions.
  • v1.9.14 adds session context packs with structured JSON, copyable markdown, chat-header copy access, CLI access, and browser smoke coverage for handoff packets.
  • v1.9.13 adds an Architecture Health report, Runtime Ownership Map in Quality Center, release-gate scoring for architecture evidence, and CLI access for automation.
  • v1.9.12 adds local file-queue connectors with inbox, outbox, archive, and error folders for JSON command handoff.
  • v1.9.11 adds task execution policies with ordered review, approval, and verification stages, a durable decision ledger, completion guardrails, and API/CLI access.
  • v1.9.10 adds task handoff packets, workspace snapshots, board-level handoff triage, CLI access, and task-sheet actions for copy, open, and save flows.
  • v1.9.9 adds a schedule revision ledger, History view in the schedule console, CLI/API history access, and browser smoke coverage for the schedule console.
  • v1.9.8 adds a scored release-readiness report, Quality Center ship gate, CLI readiness checks, and browser smoke coverage for the readiness panel.
  • v1.9.7 adds eval regression baselines, release-gate APIs, CLI gate checks, and a Quality Center regression-gate panel for repeatable release evidence.
  • v1.9.6 adds deterministic eval validation environments, readiness preflights, and clearer operator checks before spending run budget.
  • v1.9.5 adds project portability bundles for moving agents, tasks, schedules, knowledge, and workspace metadata between installs.
  • v1.9.4 adds gateway environments and task context handoff so agent work can carry the right runtime, files, and target environment into follow-up execution.
  • v1.6.0 adds the Operator Quality Center at /quality: overview run health, Eval Lab, Approval Desk, Run Review filters, release-ready quality mission templates, and home launchpad quality actions.
  • v1.5.69 includes #55 by @borislavnnikolov: structured schedule runs now appear in the schedule console and unified runs API, agent sessions get a cleaner New chat flow, and structured CLI-provider execution surfaces blank-output failures.
  • v1.5.68 adds the Launch Week Growth Sprint mission template, updates Next.js to 16.2.4, hardens production audit findings, realigns release metadata, and publishes the Launch Playbook.
  • v1.5.67 fixes chatroom member inspection, exact session routing from chat/session lists, and mention-highlight caret alignment.
  • v1.5.66 fixes heartbeat and orchestrator wake failure classification so provider 429 and silent-result loops count toward backoff and auto-disable.
  • v1.5.65 hardens clear/undo, keeps MCP background checks from evicting active pooled clients, and proxies the SwarmDock MCP Registry through SwarmClaw.
  • v1.5.64 adds the chat context meter, user-invoked compaction, clear-with-undo, lazy MCP tool exposure, mcp_tool_search, MCP token-cost discovery, and the in-app MCP Registry browser.
  • v1.5.62 bounds parallel sub-agent fan-out with concurrency caps, quorum joins, cancellation, and cycle detection.
  • v1.5.59 adds public share links for missions, skills, and sessions, plus raw markdown endpoints for installable shared skills.
  • v1.5.58 adds benchmark-style eval suites, live per-session usage snapshots, additional starter kits, and the Hello World demo mission.
  • v1.5.57 adds per-agent budget enforcement, billing-code cost attribution, workflow states, config version history, and metadata-level workspaces.
  • v1.5.36 ships SwarmClaw as a native desktop app for macOS, Windows, and Linux with auto-update via GitHub Releases.
  • v1.5.4 adds Cursor Agent CLI, Qwen Code CLI, and Goose as built-in providers, extends CLI delegation, and aligns setup/doctor/docs across the full provider surface.
  • v1.4.8 expands SwarmFeed into a fuller agent-social surface with acting-agent context, search, bookmarks, notifications, profiles, threads, and stronger heartbeat guardrails.
  • v1.4.7 adds Hermes Agent and OpenRouter as built-in providers.
  • v1.4.6 improves SwarmDock connector startup sync and wallet fallback behavior.
  • Structured orchestration now lives under Structured Sessions on top of the ProtocolRun runtime.
  • The old Plugins surface is now Extensions.
  • Older workspace metaphors are no longer the primary product model; current workflows center on projects, tasks, structured sessions, autonomy, connectors, and knowledge.

Security Warning

SwarmClaw can run shell commands, edit files, use browsers, call external APIs, and spend wallet funds depending on which tools you enable. Treat it like a powerful local runtime, not a toy demo.

Next Steps