OpenClaw: The Open-Source AI-Agent Runtime

Published 2026-07-27 · By Stratronix AI · 9 min read · Technical Series

Why OpenClaw exists: AI agent infrastructure should be auditable. Closed-source AI agent platforms create vendor lock-in and unverifiable security claims. STRATRONIX built OpenClaw as a fully open-source (Apache 2.0) AI-agent runtime so every enterprise can inspect, modify, and self-host the agent that holds their prompts, documents, and business workflows.

1. What Is OpenClaw?

OpenClaw is the AI-agent runtime that runs inside STRATRONIX STA-100 PAA. It is also available standalone as an open-source project (Apache 2.0 license) for organisations that want to deploy the same agent runtime on their own infrastructure (Linux server, container, VM) without buying STRATRONIX hardware.

Property	Value
License	Apache 2.0 (permissive, commercial-friendly)
Language	Python 3.11+ with Rust core for performance-critical paths
Minimum hardware	4-core ARM / x86, 4GB RAM, 64GB storage
OS support	Ubuntu 22.04+, Debian 12+, RHEL 9+, macOS, Windows WSL2
Container support	Docker, Podman, Kubernetes
Cloud LLM providers	OpenAI, Anthropic, Mistral, Aleph Alpha, OpenAI Azure, AWS Bedrock, Google Vertex, Aliyun, DeepSeek, Ollama, vLLM
Embedding models	OpenAI text-embedding-3, BGE-M3, E5-large, instructor-xl, custom
Vector stores	SQLite-vec, FAISS, Qdrant, Milvus, Pinecone, Weaviate
Chat clients	Feishu, WeChat Work, Slack, Teams, Telegram, WhatsApp, Discord, Email
Documentation	docs.openclaw.ai (mirrored to GitHub Pages)
Source repository	github.com/stratronix/openclaw

2. Core Architecture

OpenClaw follows a modular agent architecture inspired by the ReAct (Reasoning + Acting) pattern. The runtime has four layers:

1. Inbound layer — receives messages from chat clients (Feishu, Slack, etc.) and web UI
2. Agent core — processes the message: classifies intent, retrieves RAG context, plans actions, calls LLM
3. Tool layer — executes actions: search documents, send emails, query databases, call APIs
4. Outbound layer — formats responses and sends back to the user via the originating chat client

3. The Agent Loop (ReAct Pattern)

For each user message, OpenClaw runs the following loop:

1. Receive message from chat client
2. Pre-process — redact PII, classify intent, check permissions
3. Retrieve — query local RAG knowledge base for relevant context (vector search + keyword)
4. Plan — agent decides: direct answer, retrieve more, or call a tool
5. Act — call cloud LLM with full context; optionally call a tool (e.g., search internal CRM)
6. Verify — check response for policy violations, hallucinations, sensitive content
7. Respond — format and send back to the user
8. Log — write structured log entry to local audit database

Each step is logged locally for auditability and replay debugging.

4. RAG Engine

OpenClaw's RAG (Retrieval-Augmented Generation) engine keeps your documents on-device:

• Ingestion — PDF, DOCX, MD, HTML, TXT, EML, CSV; chunked and embedded
• Embedding — your choice of embedding model; embeddings stored on-device
• Retrieval — hybrid (vector + keyword); configurable recall/precision trade-off
• Citation — every LLM response includes source citations to the source document + paragraph
• Refresh — incremental re-indexing when documents change

Important: Document text never leaves your device. Only the embedding vectors and the LLM context window contents are sent to the cloud LLM during a query. The cloud LLM provider sees the query text + retrieved context (which you can redact before sending), but never sees your full document store.

5. Tool Layer (Integrations)

OpenClaw ships with pre-built integrations for common enterprise tools:

Category	Tools
Chat / Messaging	Feishu, WeChat Work, Slack, MS Teams, Telegram, WhatsApp, Discord
Email	SMTP/IMAP, Gmail API, Microsoft Graph, Outlook
Calendar	Google Calendar, Microsoft 365, CalDAV
Document Storage	Local filesystem, S3, Aliyun OSS, Nextcloud, SharePoint
Databases	PostgreSQL, MySQL, SQLite, MongoDB, ClickHouse
CRM	Salesforce, HubSpot, Zoho, Pipedrive (via REST)
Ticketing	Jira, Linear, Zendesk, Freshdesk
Search	Elasticsearch, Meilisearch, Typesense
Custom	Webhook + OpenAPI / OpenRPC / MCP tools

Custom tools can be added via the MCP (Model Context Protocol) standard or simple Python function decorators. Every tool call is logged locally.

6. Configuration & Deployment

A minimal OpenClaw config (YAML) for a typical European SMB:

# openclaw.yaml — minimal European SMB deployment
agent:
  name: paa-smb
  cloud_llm:
    provider: mistral
    region: eu-west-1
    api_key_secret: env:MISTRAL_API_KEY
rag:
  embedding_model: bge-m3
  vector_store: sqlite-vec
  documents_path: /srv/openclaw/documents
chat_clients:
  feishu:
    app_id: env:FEISHU_APP_ID
    app_secret: env:FEISHU_APP_SECRET
security:
  disk_encryption: aes-256
  audit_log: /var/log/openclaw/audit.jsonl
  redaction_pii: enabled
network:
  outbound_allow:
    - api.mistral.ai
    - api.openclaw.ai (OTA updates only)

7. Audit Trail & Compliance

Every agent action is logged to a local append-only JSONL file. Each log entry includes:

• Timestamp (RFC 3339 with timezone)
• User identity (chat client user ID)
• Intent classification
• Retrieved document IDs
• LLM call: provider, model, prompt hash, completion hash, token counts
• Tool calls: tool name, inputs (redacted), outputs (truncated)
• Final response (full text)
• Verification results (PII scan, hallucination score, policy check)

Audit logs can be exported to your SIEM (Splunk, Elastic, Datadog) via OpenClaw's webhook integration.

8. Security Model

OpenClaw's security model assumes the host network is hostile. Defences include:

• TLS 1.3 for all outbound calls to cloud LLMs
• PII redaction before sending prompts to cloud LLMs (regex + ML-based)
• Prompt injection defence — content from documents treated as untrusted; system prompts isolated
• Tool allowlisting — only configured tools can be called by the agent
• Rate limiting — per-user and per-API-key throttling
• Sandboxing — tool execution in containerised sandboxes (gVisor / firecracker)
• SBOM + signed releases — every release includes a Software Bill of Materials and cryptographic signatures

9. OTA Updates

STRATRONIX-managed OpenClaw instances (running on STA-100) receive automatic OTA updates for security patches and new features. Update cadence:

• Security patches: within 24 hours of upstream fix
• Feature releases: monthly, with 7-day staged rollout
• Major versions: quarterly, opt-in

Self-hosted OpenClaw installations can opt out of OTA and pin to a specific version. STRATRONIX commits to maintaining the v1.x branch for 3 years from release.

10. Self-Hosting OpenClaw (Without STRATRONIX Hardware)

If your organisation prefers to deploy OpenClaw on existing infrastructure:

1. Clone the repository — github.com/stratronix/openclaw
2. Install dependencies — pip install openclaw (or docker pull stratronix/openclaw)
3. Configure — openclaw.yaml with your cloud LLM provider and document paths
4. Run — openclaw serve --config openclaw.yaml
5. Bind chat clients — scan QR code from the admin UI

Self-hosted instances receive the same OTA updates (configurable) and access the same documentation. STRATRONIX charges no licensing fee for OpenClaw itself; revenue comes from the STA-100 hardware.

11. Roadmap (2026-2027)

• Q3 2026: v1.4 — multi-agent orchestration, parallel tool calls
• Q4 2026: v1.5 — on-device speech-to-text (Whisper-cpp) for voice input
• Q1 2027: v2.0 — typed tool definitions, formal verification of agent plans
• Q2 2027: v2.1 — federated OpenClaw instances for multi-site enterprises

12. Why Open Source?

Closed-source AI infrastructure creates unverifiable security claims. Every line of OpenClaw that runs on your device is auditable. Security researchers can inspect, modify, and self-host the runtime. This is a deliberate architectural choice: open-source is the only path to enterprise trust at scale. STRATRONIX commits to keeping OpenClaw open-source indefinitely; the business model is built around hardware and managed services, not software lock-in.