← Back to Skills Marketplace

Mulch

Name: Mulch
Author: austindixson

by austindixson · GitHub ↗ · v1.0.5

cross-platform ⚠ suspicious

694

Downloads

Stars

Active Installs

Versions

Install in OpenClaw

/install mulch

Description

Mulch Self Improver — Let your agents grow 🌱. Captures learnings with Mulch so expertise compounds across sessions. Use when: command/tool fails, user corre...

Usage Guidance

This skill appears to do what it says: it provides a local, git-tracked 'Mulch' workflow and an OpenClaw hook that injects a short reminder and suggests using the mulch CLI. Before installing or running any provided scripts, review the repository's scripts (activator.sh, error-detector.sh, extract-skill.sh and the docker-test scripts) in a sandbox or CI environment to confirm they behave as expected. Note two small mismatches to verify: (1) the project claims Telegram notifications but the code does not include an implemented integration or require Telegram credentials — if you enable notifications you should verify where/how credentials are configured; (2) small metadata inconsistencies (owner/slug differences in _meta.json vs registry metadata) likely indicate packaging/versioning noise but you may want to confirm the source/repository identity before trusting automated runs. If you plan to run the included Docker test or any scripts, run them in an isolated environment (container) and inspect the output before granting broader access.

Capability Analysis

Type: OpenClaw Skill Name: mulch Version: 1.0.5 The core 'mulch' skill is benign, but the bundle includes the 'visual-explainer' skill, whose prompt templates (`visual-explainer/prompts/diff-review.md`, `visual-explainer/prompts/fact-check.md`, `visual-explainer/prompts/plan-review.md`, `visual-explainer/prompts/project-recap.md`) instruct the AI agent to execute shell commands (`git diff`, `git log`, `grep`, `open`, `xdg-open`, `surf gemini`) using unsanitized user input (`$1`, `$2`, `$@`). This creates significant shell injection and path traversal vulnerabilities, allowing a malicious user to potentially execute arbitrary commands or read/write arbitrary files on the host system. Additionally, the use of `npx --yes mulch-cli` in benchmark scripts, while common, introduces a supply chain risk if `mulch-cli` itself were compromised. The mention of Telegram notifications also indicates a data exfiltration *capability*, though not maliciously implemented by this skill.

Capability Assessment

✓ Purpose & Capability

The skill's name/description (a self-improvement layer that records learnings with Mulch) matches the code and docs: it adds a bootstrap reminder, suggests running `mulch prime`/`mulch record`, and ships provider hooks and helper scripts. Declared capabilities (domains, search, recording) align with the included CLI usage and hook handlers.

✓ Instruction Scope

SKILL.md and the OpenClaw hook limit behavior to adding a virtual reminder file at bootstrap, suggesting `mulch` CLI commands, and prompting users to record learnings. The runtime instructions operate on local repo state (.mulch/) and do not direct the agent to read unrelated system files, environment secrets, or to contact unexpected external endpoints.

✓ Install Mechanism

There is no install spec (instruction-only skill for the platform), so nothing is downloaded or installed automatically. The README recommends installing the public mulch-cli via npm or using npx, which is proportional to the stated purpose. Included scripts and tests are local to the skill repository.

ℹ Credentials

The skill declares no required env vars or credentials (appropriate for a local CLI/helper). However, config/domains.json and multiple docs mention notifications via Telegram (notifications.enabled: true and 'You'll be notified via Telegram') even though no Telegram token/chat-id env var is declared or used in the hook code (notifyUser is a stub that does not perform network calls). This is a mismatch to be aware of: notifications are advertised but there is no built-in credential/config wiring in the shipped code.

✓ Persistence & Privilege

The skill does not request always:true and does not modify other skills' configurations. The OpenClaw hook injects virtual files at bootstrap (SELF_IMPROVEMENT_REMINDER.md, MULCH_PRESET_DOMAINS.md) which is expected behavior for a reminder hook. Autonomous invocation of the hook on bootstrap is the normal platform pattern.

How to Use

Make sure OpenClaw is installed (local or Docker)
Run the install command in chat: /install mulch
After installation, invoke the skill by name or use /mulch
Provide required inputs per the skill's parameter spec and get structured output

Version History

v1.0.5

No user-facing changes in v1.0.5. - No file changes detected in this release. - Documentation and functionality remain unchanged.

v1.0.4

Mulch v1.0.4 introduces auto-detection and notification features, plus expanded documentation and domain support. - Automatic detection of errors and user corrections, prompting to record new learnings. - Notifications sent via Telegram when a learning is recorded. - 24 preset domains included for easy setup (see `config/domains.json`). - Extensive documentation, references, templates, and visual explainers added. - SKILL.md updated with streamlined instructions and feature overview.

v1.0.3

**Clarifies Mulch’s role as a passive layer in intro and description.** - Description and introduction now clearly state: "Mulch is a passive layer. It does not contain an LLM. Agents use Mulch — Mulch does not use agents." - No functional or workflow changes; content and guidance remain the same. - No file changes detected—documentation update only.

v1.0.2

**Changelog for mulch v1.0.2** - Clarified and emphasized Mulch's token and character efficiency (~54% reduction) in descriptions and intro. - Added explicit benchmark data callouts and references to BENCHMARK.md in main documentation. - No functional or file changes; documentation improved for clarity and more concise communication of benefits. - Maintained all prior usage guides, examples, and setup steps.

v1.0.1

- Added benchmark and qualification documentation (BENCHMARK.md, QUALIFICATION.md) to provide transparency on performance and feature coverage. - Minor metadata and documentation updates; no changes to core functionality. - No breaking changes; existing workflows and usage remain the same.

v1.0.0

Mulch Self Improver skill initial release: - Enables agents to persist learnings across sessions using Mulch (`mulch record`, `mulch query`, `mulch prime`). - Guides agents to record failures, user corrections, missing features, and better approaches for compounding expertise. - Provides structured workflows, record types, and promotion pathways for best practices and recurring patterns. - Integrates with popular agent platforms (Cursor, Claude, Codex, Gemini, etc.) via setup hooks. - Supports regular review, pruning, and promotion of learnings to project or workspace memory files (AGENTS.md, CLAUDE.md, TOOLS.md, SOUL.md).

Metadata

Slug mulch

Version 1.0.5

License —

All-time Installs 4

Active Installs 4

Total Versions 6

Frequently Asked Questions

What is Mulch?

Mulch Self Improver — Let your agents grow 🌱. Captures learnings with Mulch so expertise compounds across sessions. Use when: command/tool fails, user corre... It is an AI Agent Skill for Claude Code / OpenClaw, with 694 downloads so far.

How do I install Mulch?

Run "/install mulch" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Mulch free?

Yes, Mulch is completely free (open-source). You can download, install and use it at no cost.

Which platforms does Mulch support?

Mulch is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Mulch?

It is built and maintained by austindixson (@austindixson); the current version is v1.0.5.

More Skills