Habr AI→ original

Domclick Developer Taught OpenCode to Configure Itself: MCP, Skills and Permissions in a Couple of Evenings

Domclick backend developer Vladimir Bandov has published a practical write-up on Habr about configuring a coding agent in OpenCode Desktop. The first half hour of setup went smoothly. Then he got buried in skills, MCP servers, access permissions and several config files. The agent also took unnecessary steps and implemented things he had not asked for. After a couple of evenings, the author settled on a "lazy" setup: the agent helps configure itself, while the human stays in charge. The price of the approach is possible token overspending. Here, the agent is an assistant, not a full-fledged developer. The plan must come before access permissions and terminal commands. The author reduced adding MCP servers and skills to a minimum of manual steps.

AI-processed from Habr AI; edited by Hamidun News
Domclick Developer Taught OpenCode to Configure Itself: MCP, Skills and Permissions in a Couple of Evenings
Source: Habr AI. Collage: Hamidun News.
◐ Listen to article

Domclick backend developer Vladimir Bandov has published a write-up on Habr about configuring a coding agent in OpenCode Desktop. The first half hour of setup went without problems. Then came chaos: skills, MCP servers, permissions and several settings files. Finding a workable approach took a couple of free evenings. The author calls the result "lazy": it may spend more tokens, but the agent helps configure itself and does not go off on its own.

Where the setup breaks down

According to Bandov, the barrier to entry turned out to be higher than the tutorial videos suggest. You can install the agent in a few minutes. Making it work predictably is harder. The author names three problems.

First: too many entities. Skills, MCP, permissions and different configuration levels all have to be kept in mind at once. At some point he stopped remembering which skill was for what.

Second: the order of actions. He had to explain to the agent over and over that it first needs a plan, then access permissions and configurations, and only then can it run commands in the terminal.

Third: hyperinitiative. The agent took unnecessary steps and invented and implemented things that were not part of the task. For a working repository, this is the main risk: you have to read and roll back the extra code.

The lazy approach: the agent configures itself

The idea from the article's title is to manage the configuration "through the agent's own hands." Instead of editing files manually and remembering the syntax, the developer sets a task, and the agent proposes a plan and makes the changes: it connects MCP, adds skills and writes permissions. The author walks through the full steps and prompts in the original; the announcement does not include them.

Some context for those who haven't followed this. MCP (Model Context Protocol) is an open protocol that Anthropic introduced in November 2024 for connecting external tools and data sources to models. OpenCode is an open-source coding agent that runs in the terminal and in a desktop client. Its permissions for actions are set in configuration using an "allow, ask, deny" scheme. Once there are several MCP servers and each has dozens of tools, manual administration quickly turns into a chore.

The framing of the use case matters too. Bandov does not delegate the role of a full-fledged developer to the agent. For him, the agent is an assistant: it prepares a plan, carries out clear steps and asks where control is needed. This mode also explains his willingness to pay in tokens for extra planning iterations.

What the announcement tells us

  • Tool: OpenCode Desktop.
  • Author: Vladimir Bandov, backend developer at Domclick.
  • Time: about 30 minutes until manual setup became tiring; a couple more evenings to arrive at a working scheme.
  • Components: skills, MCP, permissions, several configurations.
  • Main problem: the agent does work it was not asked to do.
  • Cost of the solution: possibly more tokens spent.
  • The announcement contains no measurements of time or cost savings.

How to apply this yourself

The piece offers several practical rules that do not depend on a specific client.

Start with a plan. If the agent must first describe its steps and only then get access to commands, write that into the project rules so you don't have to repeat it in every conversation.

It is better to set permissions strict by default. You can allow reading files, while writes and shell commands should require confirmation. Loosen restrictions selectively, for a specific task.

Connect MCP servers one at a time and check which tools each adds to the context. Every server spends tokens on descriptions and widens the surface for errors.

Keep the configs the agent edits in git and review them via diff. If the agent changes its own permissions, you need to see that before the commit.

Finally, measure. If the "lazy" setup really costs more tokens, evaluate that on real tasks over a week and compare it with the time you no longer spend on manual configuration.

ZK
Hamidun News
AI news without noise. Daily editorial selection from 50+ sources. A product by Zhemal Khamidun, Head of AI at Alpina Digital.
✦ Autonomous AI Agents

Delegate Deep Research, Coding & Workflows to an AI Agent

Deploy autonomous multi-step execution with NomadNet Agent (agent.nomadnet.ai) — automated web research, data synthesis, code generation, and complex task orchestration.

Launch NomadNet Agent →NomadNet AI Platform
What do you think?
Loading comments…