Valmis: An Open-Source, Security-Focused AI Agent Framework

Valmis is an open-source cloud-based application designed to deploy AI agents for production environments. Unlike many personal assistant tools, Valmis prioritizes enterprise-grade security by ensuring that AI agents never have direct access to API credentials or host files, instead interacting with business tools through a secure proxy system.

Secure Architecture via Proxy System and Isolation

Valmis solves the security risks associated with agents storing credentials in plain text or sending them to LLM providers by implementing a proxy-based architecture.

Credential Isolation: AI agents run in isolated Docker containers. They cannot access API keys directly; instead, they request the host machine to make API requests by providing a credential ID. The host machine executes the request and returns only the JSON data to the agent.
Network Control: Because the host handles all API and LLM calls via the proxy, the agent container's internet access can be completely disabled without breaking its functionality.
File System Isolation: Each agent possesses its own dedicated file system, remaining entirely isolated from the host machine and other agents.
Encryption: Credentials are encrypted using AES-256-GCM and stored securely in the database.

Agent Capabilities and Workflow Automation

Valmis allows users to build a fleet of independent or collaborative agents, each with its own LLM provider, knowledge base, and set of credentials.

Multi-Step Workflow Automation

Users can automate repetitive tasks using a workflow builder canvas. These workflows can be triggered by cron jobs, webhooks, or specific app events (such as a new email or form submission). The system supports:

Smart and Strict Conditions: Conditions can be defined using natural language (AI-decided) or rigorous programming logic.
Loops and Data Mapping: Workflows support loops and defined output schemas for efficient data mapping between steps.
Granular Permissions: Users can limit which credentials and tools are available to an agent at each specific step of a workflow.

Browser Automation

Agents can operate a headless browser to navigate websites, fill out forms, click elements, and take screenshots. To maintain security, these browsers are managed by the host machine, and agents interact with them via the proxy. Agents also maintain their own browsing history and session cookies.

Cognitive Memory and Knowledge Management

Valmis implements a memory system inspired by cognitive research, utilizing pgvector for semantic search and text embedding.

Cross-Session Memory

Agent memory is persistent across sessions and categorized into four types:

Episodic: Records of what happened.
Semantic: Durable facts.
Procedural: Rules and constraints.
Working: Short-lived context.

Agents automatically distill learned information at the end of a session to improve future interactions. Users can also manually instruct agents to modify or remove specific memory items.

Knowledge Base Integration

Users can connect enterprise knowledge bases via Google Drive, Dropbox, and Notion, or by uploading files. These files are processed as memories to ensure rapid knowledge recall.

Integrations and Extensibility

Valmis supports over 100 business and productivity integrations, including Google Workspace, Slack, Notion, HubSpot, Salesforce, and Figma.

Extensibility: Every integration is defined by a single YAML file, making the catalog easy to extend.
LLM Flexibility: The platform supports nearly 200 models from 20 providers, including OpenAI, Anthropic, Google, Mistral, and Cohere, as well as OpenRouter.
Skills System: Users can install third-party skills from GitHub or create self-evolving skills that learn through interaction.

Human-in-the-Loop and Specialized Tools

To prevent autonomous errors in critical tasks, Valmis includes a "Human in the loop" feature where the agent pauses and presents the human user with a set of options before proceeding with a critical decision.

As a demonstration of its tool-use capabilities, Valmis includes a chess-engine tool. Rather than relying on the LLM to hallucinate moves via text generation, the agent uses a lightweight built-in chess engine to calculate legitimate moves, allowing the AI to play valid chess games.

Deployment

Valmis is released under the Apache License 2.0 and can be deployed using Docker Compose, which orchestrates the frontend, backend, a pgvector-enabled PostgreSQL database, and a Docker socket proxy for the agent runtime.