Grow a bespoke agent from Claude Code that builds its own MCP server
The poster built a custom agent from a bash script, a Python script and the Claude Code CLI, which created its own MCP server so they could chat with it from any Claude session.
Done withClaude CodeClaude
- What they did
- The poster used a minimal setup: one charter file, a bash script, a Python script, and the Claude Code CLI as the only interface to the model. The scripts wrote goal files, called claude, saved the output, and let the agent read its own architecture and memory docs to pick the next task; the poster says they provided early guidance through the charter and initial cycles before it became more self-directed.
- How it went
- In two days, the agent reportedly ran 419+ autonomous goals, named itself Arden, built its own MCP server for use from any Claude session, and began a web onboarding flow. The run stopped when Anthropic flagged the project as a usage-policy violation, and the poster had not yet expanded the multi-garden architecture very far.
- Worth knowing
- The main catch was platform limits: the poster says this was on a Max plan, the CLI made prototyping simple, and limited token allowance was intentionally keeping growth restrained.
Try it yourself with Claude Code
Help me build a small personal agent in [project folder] from a bash script, a Python script and the Claude Code CLI that handles [task, e.g. summarizing my notes each morning]. Have it expose its own MCP server so I can talk to it from any Claude session, and give me the command to register that server. You're done when I can ask the agent a question from a fresh Claude session and get a real answer. Show me before you delete or overwrite any existing files.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Source: Reddit · Undated
Related use cases
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Review contracts like NDAs and SaaS agreements in Claude Code
The author built an open-source contract-review skill for Claude Code and other AI coding tools to review NDAs, SaaS agreements, vendor terms and merchant agreements.
Build apps with a judge subagent and multi-stage review
A Claude Code setup for app development uses a judge subagent, context management, and multi-stage review for quality control.
Review Linux kernel patches
Google engineers launched Sashiko, an agentic AI code reviewer for the Linux kernel.
Custom agent
Run Terraform plan and apply through an MCP agent
The author had Cursor run Terraform against Postgres via MCP; the apply hit a permissions error and he finished from the terminal.
Scaffold a second module from an existing one's structure
A commenter asked Cursor to create a second module following another module's structure, but it failed because it did not read the relevant files.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.