Autonomous Digital Workforce

GoClaw is a high-performance Multi-Agent Gateway and Cognitive Operating System compiled into a single binary. Seamlessly coordinate LLM agents to execute strict workflows.

📖 Documentation ↗

🧠

Cognitive Architecture & Memory

GoClaw moves beyond simple chatbot loops. Every interaction updates a structured facts database, enabling long-term personalization and learning.

Knowledge Graph RAG: Hybrid relational + vector mapping.
Continuous Fact Extraction: Auto-extracts configurations and user rules.
Memory Pruning: Automatically consolidates, deduplicates, and prunes stale memories.
Any Provider: Gemini, OpenAI, Claude, DeepSeek, Ollama, or any OpenAI-compatible endpoint — with automatic failover.

🔌

Omnichannel & Secure Ingress

Orchestrate specialized personas across chat platforms and your own apps, each with its own identity and access rules.

Built-In Channels: Discord, Telegram, MS Teams, and Facebook Messenger. Discord and Telegram need no public URL.
Enterprise App API: Drop agents into your own mobile app or intranet over REST with SSE streaming, with per-employee isolated memory.
Safe Webhook Router: Every route checks an HMAC signature or token before it runs a SOP; the agent can set up the hook in GitLab, SonarQube, Graylog, Grafana, Zabbix, Uptrace, and Wazuh.
AES-256 Vault: Safe credential lookup via encrypted secrets.

🛠️

120+ Built-In Sandboxed Tools

All core capabilities are compiled into the one Go binary (SQLite and tree-sitter are linked in through CGO). No external interpreters or plugins to install.

AST Code Analysis: Tree-sitter powered syntax parsing and refactoring.
Database Queries: Read-only queries on PostgreSQL, MySQL/MariaDB, ClickHouse, and MongoDB; writes only where a datasource allows them.
Desktop & Web: Sandboxed headless browser actions and shell constraints.

⚡

Event-Driven Proactivity

GoClaw implements a concurrent worker pool and async actor models. Agents proactively start diagnostics based on alerts.

Non-Blocking Event Streams: ProcessStream interface for fluid client output.
SOP Automation: Execute multi-step tasks with branch fallbacks and retry limits.
WorkerPool Concurrency: Bounded task queue preventing memory overflows.

The 5 Levels of AI Autonomy

GoClaw bridges the gap between simple chatbots and fully autonomous Level 4 digital employees. We design systems that operate within deterministic, safe corporate guardrails.

Autonomy Level	GoClaw Implementation	Target Capabilities	Guardrails & Control
L1: Chatbots	100% Supported CLI Chat & Personality	Conversational Q&A based on static prompt context. No external actions.	Strictly conversational. No access to tools or environment.
L2: Copilots	100% Supported Codebase AST Indexing	Reads local environment, workspace files, or active git context to suggest solutions.	Read-only access. Humans must copy/apply suggestions manually.
L3: Autonomous Agents	100% Supported ReAct Loops & Webhooks	Uses ReAct loops to pick tools and fix issues dynamically based on general prompts.	Higher risk. Prone to loops, environment pollution, and rate limits.
L4: AI Employees	95% Supported Production-Ready SOPs	Follows strict SOPs, manages persistent knowledge base, and collaborates asynchronously.	Per-step compensating rollback commands, an effects journal for safe resume, loop protection, and human-in-the-loop approvals.
L5: Self-Directing	Not Implemented Vision Only	Agents propose goals from operational signals, test hypotheses, and suggest SOPs and tools.	Only as governed autonomy: proposals go through human review and policy checks. Unsupervised self-direction stays out of scope.

🛡️ Security Philosophy

GoClaw treats the agent not as an all-powerful script runner, but as a sandboxed employee. Core guidelines enforce maximum security:

Built-In Over Bolted-On: Single Go binary. No dynamic library runs.
Jailed Workspaces: File reads and writes are jailed to the workspace folder.
Compensating Rollbacks: A SOP step can declare a rollback_command; when a later step fails the run, the passed steps' commands run newest first.
SSRF Protected Clients: Safely dials third-party APIs without loopbacks.

💡 Self-Correction & Learning

When a tool call fails on a wrong table, column, action or instance, GoClaw returns the error with a hint and the matching approved Saved Queries or Tool Presets, so the agent corrects the call in the same turn. Nothing is stored from the failure: what the agent should do every time is written down by an admin as a saved query, a tool preset or a wiki page.

⚡ Quick Installation Script

curl -fsSL https://goclaw.ulab.mn/install.sh | sh

Downloads the appropriate platform binary, verifies its SHA-256 checksum, and configures path aliases.

Build From Source

# Clone repository and build the gateway binary
go build -o goclaw ./cmd/gateway/

# Launch the interactive configuration wizard
./goclaw setup

Core CLI Commands

goclaw setup                            # Interactive configuration wizard
goclaw                                  # Start the HTTP/WebSocket gateway
goclaw chat                             # Terminal chat session (no gateway needed)
goclaw status                           # Runtime and connectivity diagnostics
goclaw memory stats                     # Active RAG database statistics
goclaw tools list                       # Every registered tool
goclaw service install                  # Desktop: per-user background service
sudo goclaw service install --system    # Server: system service (runs as your user)
goclaw service restart                  # Restart the installed service
goclaw update                           # Update binary and restart the service

State Directory Layout

~/.goclaw/
├── config.yaml          # Hot-reloaded YAML configuration
├── gc_state.sqlite      # Sessions, memory, SOPs, cron jobs
├── kg.sqlite            # Knowledge graph store
├── workspace/           # IDENTITY.md, USER.md, Memory.md
├── agents/              # Isolated per-agent environments
│   └── <agent_id>/
│       └── workspace/   # Agent-scoped files and memory
├── memory/personal/     # Per-user private memory (user_<id>.md)
├── credentials/         # AES-256 encrypted secrets vault
├── skills/              # Skill text context files
├── plugins/             # Installed plugins
├── media/  canvas/      # Uploads and generated artifacts
├── traces/archive/      # Nightly trace dumps (JSONL.gz)
└── logs/                # System OTel diagnostics

⚠️ Safety & Responsibility Notice: GoClaw is developer/admin software. You are responsible for how you configure, provision, and authenticate agents, connected databases, credentials, and cron tasks. Binary usage is permitted under the proprietary GoClaw Binary License.