Policy before action
Each tool call declares what it intends — file write, process exec, network connect — and the policy engine answers allow, ask, or deny before any side effect exists.
MyCoder · Developer tools
A coding agent that asks before it acts.
Point MyCoder at a repository and describe a task. It reads, edits, runs your tests, and checks its own work — and every tool call is decided by a policy engine before the tool does anything.
Install
Running in under a minute.
# npm (Node 22.18+)
$ npm install -g mycoder-cli@alpha
# or Homebrew
$ brew tap OIerYangJZ/mycoder
$ brew trust OIerYangJZ/mycoder
$ brew install mycoder
# add a model provider key (read from stdin, never echoed)
$ printf %s "$API_KEY" | mycoder setup-credential ~/.config/mycoder/secrets/provider.key
# verify, then start
$ mycoder doctor
$ mycoder
Features
Built for work you can audit.
Four approval modes
Plan, manual, accept-edits, and auto. Secrets, network, VCS mutation, and MCP calls still ask in every mode, including auto.
Honest sandbox reporting
Enforcement is reported per dimension — none, best-effort, policy, container, or OS — so you know what is actually isolated.
Undo with receipts
Atomic writes with a unified diff and rollback metadata. /undo reverses an edit, a turn, or a file, and says what it could not cover.
Bring your own model
Adapters for Anthropic Messages, OpenAI Responses, and OpenAI-compatible APIs such as DeepSeek, OpenRouter, or a local llama.cpp.
Scriptable
--json emits one object per line on stdout, so MyCoder drops into CI and shell pipelines without filtering prose.
Support
Found a bug or need a feature?
Open an issue
Or email contact@jinseylab.tech