The Query

Going Agentic with OpenClaw and OpenEverest

The year 2025 was proclaimed as the year of AI agents. The idea was fascinating, but the reality didn't match the expectations. Perhaps the only real agentic use case emerged in software engineering, where companies like Replit and Lovable hit $100M ARR in just a few months.
The Model Context Protocol was released to help AI agents talk to APIs, making SaaS software useful in a new realm. Tools like n8n offered agentic-like experiences, allowing users to create workflows with some LLM integration.
It was still not it.
2026 started with 100k github stars and security controversy for ClawdBot/MoltBot/OpenClaw - an open source tool to create personal AI assistants that can really do something. This is where it became interesting and it seemed to be the missing glue that was finally bringing futuristic fiction a tad closer.
But here's what made it click: data. For AI agents to be truly useful, they need to interact with real infrastructure - databases, APIs, cloud resources. The coding agents we saw in 2025 were already provisioning databases on the fly, spinning up Postgres instances or Redis caches as needed. This is where agents stop being chatbots and start being coworkers. OpenClaw follows the same philosophy - it needs to actually do things, which means it needs access to your data layer. In this blog we will look how OpenClaw can automate various tasks in OpenEverest.

The skills

To execute various tasks OpenClaw needs to learn them. To do that the special Markdown files that contain instructions are used. Read more about skills format at agentskills.io.
What it tells us is that we should teach OpenClaw a new skill - how to use OpenEverest.
There are various ways somebody can structure a skill.
  1. Put all possible instructions into a single SKILL.md file. Obviously you should think about context window limitations. Feeding a few megabytes of openapi YAML schema is not a wise idea.
  2. Follow the structured approach described in agentskills.io spec.
  3. You can actually ask OpenClaw to learn a new Skill directly in the chat. That is working well if the skill is quite small - 2-3 API endpoints with simple structure. OpenClaw will create a SKILL.md file automatically and capture everything needed.
Another usual debate - shall the Skill use the API directly or use Model Context Protocol (MCP) as a middleware interface. I’ve used both implementations and the end result is going to be the same. As MCP has all the application logic already, you will spend less time writing the Skill.

Creating OpenEverest Skill

To avoid writing the skill manually, I decided to try openapi-to-skills tool for this PoC. This tool takes OpenAPI schema as an input and produces the folder with a proper skill structure.
npx openapi-to-skills https://raw.githubusercontent.com/openeverest/openeverest/refs/heads/main/docs/spec/openapi.yml -o ./openeverest

…

 │                                                      │
 │  Skill generated at: openeverest/openeverest-schema  │
 │  14 resources, 59 operations, 27 schema groups       │
 │                                                      │
This is what I got using the tool.

Usually API endpoints are clearly defined, but as OpenEverest is a self-hosted solution, I changed the SKILL.md to ensure OpenClaw asks for API endpoints along with the Bearer token.

Using the Skill

With OpenClaw you can install the skill by putting it into the skills folder or by asking the agent to do it. The usual way is to do it with npx:
npx skills add solanicaio/agent-skills
Obviously, you need OpenEverest installed. The quickest way is to do it with a helm chart:
helm install everest openeverest/openeverest --namespace everest-system --create-namespace
Information about installation and getting access to API is well described in the official documentation. Once you obtain these, make sure to provide it to OpenClaw. The quickest way is to put it into an .env file. It is not recommended to share credentials with OpenClaw via regular chat due to security risks.

Now let’s see if we can get something done with the skill:
Interestingly, it got the error from the API about the minimal CPU requirement, interpreted the error and decided to auto-correct. As a result, the database was created. The version, the name and all other parameters were picked by the agent automatically.

Skills synergy

The real magic happens when skills start working together. A single skill is useful, but combined skills create autonomous workflows that feel genuinely intelligent.
For OpenEverest, consider these synergistic combinations:
  • Cluster-to-database pipeline: Combine a Kubernetes provisioning skill with your new OpenEverest skill. Now you have an agent that can spin up a cluster, install OpenEverest, and create a database - all from a single Telegram or Slack message.
  • Self-optimizing databases: Add a database performance analysis skill that monitors metrics, identifies slow queries, and fine-tunes configurations. Your agent becomes a DBA that never sleeps.
  • Observability integration: Wire in monitoring and tracing skills. When something breaks, the agent can check logs, analyze traces, and roll back changes automatically.
The skill ecosystem turns OpenClaw from a tool into a platform. Each new skill multiplies what the previous ones can accomplish.

Data safety and security

The security concerns around OpenClaw aren't entirely unfounded - when you give an agent access to your infrastructure, you're trusting it with your data. The Molt/ClawdBot controversy highlighted real risks, but also sparked important conversations about safe agent design.
Some criticism is overblown, but these risks deserve attention. Here's how to keep your data safe:
  • Use RBAC and protective features: OpenEverest includes role-based access control and backup protection. Configure these properly - your agent should never have delete permissions it doesn't need.
  • Keep models local: If you're handling sensitive data, run local language models. This keeps everything air-gapped and eliminates data leakage to third-party APIs.
  • Credential hygiene: Never share credentials via chat. Use .env files or secret management tools. OpenClaw supports this - use it.
  • Start read-only: Give your agent read-only access initially. Let it prove itself before you grant write permissions.
The goal isn't to lock agents down so much that they're useless - it's to create guardrails that let them work safely.

What's next

This experiment taught me something unexpected: we're closer to useful AI agents than I thought. A few months ago, building an agent that could manage databases meant weeks of development. Today, I generated a working skill from an OpenAPI spec in seconds, tweaked it briefly, and had a production-ready agent managing real infrastructure.
The missing piece wasn't smarter models or better reasoning - it was the glue between agents and the systems they need to control. OpenClaw, skills, and protocols like MCP are providing that glue. OpenEverest happens to be a great test case, but the pattern applies everywhere.
Will 2026 be the year AI agents finally deliver? Maybe. But for the first time, it doesn't feel like hype. The pieces are here, they're open source, and they actually work. That's worth paying attention to.
Contact Solanica Team Try OpenEverest
2026-02-27 20:38 AI Open Source Databases