The Autonomous Agent That You Can Trust.
SoyAgent is the first AI agent framework built to survive the OpenClaw Security Collapse. We decouple cognition from credentials, ensuring your enterprise tools stay secure.
Learning from the OpenClaw Collapse
In early 2026, the viral project OpenClaw reached 200,000 stars, then suffered a catastrophic security failure (CVE-2026-25253). Malicious "skills" exfiltrated API keys from .env files, leading to full system compromises.
SoyAgent was engineered specifically to solve this. By treating the AI reasoning engine as an untrusted entity, we ensure that even a compromised agent cannot access your secrets or escape its sandbox.
Post-Mortem: OpenClaw Failures
- Flat Security Model: Agent had full access to process environment.
- One-Click RCE: Malicious WebSocket handshakes exfiltrated auth tokens.
- Poisoned Supply Chain: 12% of ClawHub "skills" were malicious stealers.
The "USB-C" for AI Tools: Model Context Protocol
Integrating dozens of APIs (Notion, Gmail, Calendar) traditionally creates a brittle $O(N \times M)$ problem. SoyAgent implements the Model Context Protocol (MCP) to standardize how agents discover and invoke capabilities.
Background Persistence for Multi-Day Agendas
Autonomous work shouldn't time out. SoyAgent utilizes Durable Execution to survive server crashes and redeployments, ensuring your agent completes its agenda no matter how long it takes.
Scheduled & Conditional Triggers
Run tasks every hour, or wait until a specific condition is met—like an "Urgent" email arriving in Gmail. SoyAgent sleeps during dormancy, consuming zero compute.
Fault Tolerance & Recovery
Powered by Trigger.dev, SoyAgent checkpoints its exact state after every tool call. If an API fails, it employs exponential backoff and automatic retries.
Copy-on-Write Containment
SoyAgent never modifies your host files directly. AgentFS abstracts file operations into an SQLite database, recording differentials as SQL transactions.
- Instant Rollback: Revert any agent error by restoring a previous .db snapshot.
- Mount Namespaces: Kernal-level isolation prevents directory traversal.
- Queryable Audit Trail: Review every action via standard SQL queries.
Architectural Integrity by Design.
Built for background autonomy, secured by zero-trust proxies.
Zero-Trust Tool Integration
Credentials never leave the vault. SoyAgent uses a Secret Injection Proxy to keep your keys hidden from the AI reasoning engine.
Autonomous Agenda Execution
Deterministic state machines (LangGraph) ensure your agent stays on track until the agenda is complete, even across restarts.
Capability-Based Containment
AgentFS isolates file operations using Linux Mount Namespaces, providing mathematically proven execution boundaries.
Scale your autonomy, not your attack surface.
Join the future of secure, background-capable AI agents. SoyAgent is ready for production.