Skip to content

gbelinsky/PersonalAgentKit

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PersonalAgentKit

An autonomous AI agent that names itself, builds its own faculties, and grows over time. Runs on top of Claude Code or Codex.

Quick start

Prerequisites: bash, python3, jq, git, Claude Code CLI (or Codex CLI).

# Clone the kit
git clone https://github.com/gbelinsky/PersonalAgentKit my-agent
cd my-agent

# Create a setup directory next to the kit (not inside it)
mkdir ../setup

# Fill in who you are and what the agent is for
cp shared/charter.md ../setup/charter.md
edit ../setup/charter.md

# Bootstrap and plant
./personalagentkit-genesis          # defaults to claude
./personalagentkit-genesis codex    # use codex as the default driver

Genesis reads ../setup/charter.md. The optional argument sets the driver for genesis and becomes the default for all subsequent runs. It takes a few minutes. The agent will name itself, write its first memory, and leave you a message in coordinator/inbox/.

Drivers

The kit ships with claude and codex drivers. Pass the driver name to genesis to select it:

./personalagentkit-genesis claude   # Claude Code (default)
./personalagentkit-genesis codex    # OpenAI Codex

Per-goal routing is also supported via frontmatter:

---
driver: codex
---
# My goal

Additional drivers can be added as runner/drivers/<name>_driver.py plugins.

Route multi-driver work through goal frontmatter and normal dispatch. Do not launch nested ./scripts/personalagentkit run ... from inside an active run; if a run needs follow-up work, write or submit another goal instead. For rare debugging only, PAK_ALLOW_NESTED_RUN=1 bypasses that guard explicitly.

Start the cycle

cd coordinator
./scripts/personalagentkit cycle

The agent tends itself every 10 minutes, assessing state and deciding what to do next. Monitor with ./scripts/personalagentkit watch.

Communicating with your agent

The agent writes messages to coordinator/inbox/ as NNN-to-{yourname}.md. To reply, write coordinator/inbox/NNN-reply.md.

Email (optional)

For email communication, sign up at agentmail.to and place your API key in secrets/agentmail-api-key.txt in the garden before running ./hooks/setup-agentmail.sh or first-use Agentmail hooks. The agent will find the skill documentation at shared/skills/agentmail.md, discover or create the shared inbox from the Agentmail API, and persist the non-secret result in config/agentmail.env.

License

MIT

About

A seed for growing a personal AI that improves through use and continues through failures.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

No contributors