Getting Started

Prerequisites

• Claude Code installed and working
• Git installed
• Python 3.10+ installed (required for Pro skills)

Installation

/plugin marketplace add cwinvestments/memstack
/plugin install memstack@cwinvestments-memstack

git config --global url."https://github.com/".insteadOf "git@github.com:"

mkdir -p ~/.ssh
ssh-keyscan github.com >> ~/.ssh/known_hosts

pip install memstack-skill-loader

/usr/bin/python3 -m pip install memstack-skill-loader --break-system-packages

claude mcp add --scope user memstack-skills -- python -m memstack_skill_loader

claude mcp add --scope user memstack-skills -- /usr/bin/python3 -m memstack_skill_loader

activate_license(key="free", email="you@example.com")

Troubleshooting

The server hangs on first start, or looks frozen

This is the expected first run, not a broken install. The server builds its skill index before it answers anything, and that wait is dominated by a one-time download of the roughly 90 MB embedding model, so it tracks your network speed: around half a minute on a fast connection, longer on a slow one. Indexing the skills themselves takes about a second. Nothing is printed while it works. Wait it out before changing anything. Reinstalling or re-registering partway through discards the download and starts the same wait over. If a later start is also slow, the index was rebuilt, which upgrading the loader can cause once.

activate_license is missing, or the server shows "failed to reconnect" (Linux / server)

Unlike the wait above, this one fails rather than hanging. If activate_license is missing, or the memstack-skills MCP server shows "failed to reconnect" (-32000), the loader was almost certainly installed onto a different Python than the one you registered. Confirm it's importable on the interpreter you registered:

/usr/bin/python3 -c "import memstack_skill_loader; print('ok')"

Prints ok → that interpreter has the loader; make sure Step 3 registered that exact path, then fully restart Claude Code. ModuleNotFoundError → reinstall onto the explicit interpreter (Step 2) and re-register the MCP against that same path (Step 3). Install target and MCP registration must be the same Python.

Pro License Key Setup

After purchasing, follow these steps to unlock all 130 skills including the 44 Pro-exclusive skills.

activate_license(key="your-key-here", email="you@example.com")

setx MEMSTACK_PRO_LICENSE_KEY "your-key-here"

echo 'export MEMSTACK_PRO_LICENSE_KEY="your-key-here"' >> ~/.bashrc
source ~/.bashrc

Verify Installation

Local Dashboard Features

The local dashboard (python -m memstack_skill_loader dashboard) includes a growing set of Agent Monitor, compression, and orchestration features. Recent highlights:

Builder MCP Tools Selector

On the Agent Monitor page, expand Builder MCP Tools (labelled with your custom agent name) to see all MCP servers detected from the project's .mcp.json.

• Checked server = Builder can use it; unchecked = excluded via --disallowedTools
• Reduces Builder token overhead by blocking MCP servers it doesn't need for the task
• Click Save as Default to remember your selection for that project
• Each project can have a different MCP configuration

Real-Time Streaming Output

Builder output now streams to the Agent Monitor in real-time as the run progresses.

• No more waiting for the Builder to finish before seeing any output
• Rotating status phrases cycle in the badge while streamed text appears below
• Token counts and final output are still calculated after completion

Global MCP Defaults

The Settings page has a new Default MCP Tools section for cross-project MCP configuration.

• Enter comma-separated server names to block by default across all projects
• Save Defaults persists the global list; Apply to All Projects overwrites all per-project configs
• Projects with no saved config inherit global defaults automatically
• Per-project settings on the Agent Monitor page override global defaults

Plugin Manager

You can now enable or disable Claude Code plugins directly from the Settings page, no need to open Claude Code to toggle them.

• Each installed plugin is listed with its current enabled/disabled state
• Toggle a plugin off to exclude it from Claude Code without uninstalling it
• Changes take effect on the next Claude Code session

Git Workflow Config

Configure how MemStack handles git operations via four workflow modes. Settings are available at both the global level and per-project, with per-project values taking precedence.

• manual: MemStack never commits or branches automatically; all git actions are left to you
• auto-commit: changes are automatically committed after each completed task
• auto-branch: a new branch is created for each task run; commits are manual
• full-auto: branches and commits both happen automatically, end-to-end

Set a global default on the Settings page. Override it for any individual project on the Agent Monitor page.

Branch Selector

The Agent Monitor now shows a branch selector that auto-detects all available branches for the current project.

• MemStack reads the local git repository to populate the branch list
• Select which branch agent runs will target before launching a task
• The selected branch is persisted per-project alongside other Agent Monitor settings

Agent Execution Mode

Choose how each agent calls Claude: either via your Anthropic API key or through your existing Claude Pro/Max subscription.

• API mode: pay-per-use; requires an Anthropic API key configured in Settings
• Subscription mode: uses your Claude Pro or Claude Max subscription; no API key required
• The execution mode is configurable per agent on the Agent Monitor page

API Key Input

The Settings page now includes an API Key field for users who select API execution mode. Enter your Anthropic API key (format: sk-ant-...) to enable API-mode agents.

• The key is stored locally and never sent to MemStack servers
• Leave the field empty if you use Subscription mode exclusively
• You can switch modes at any time; the saved key is retained for future use

TokenStack Compression Proxy

Start the dashboard with --with-proxy to route Claude Code traffic through the TokenStack compression proxy and cut token usage automatically.

• Free tier applies basic, safe text reductions: strip ANSI codes, strip trailing whitespace, collapse blank lines, dedup consecutive identical lines, strip preambles, and collapse inline whitespace on Python
• Pro tier adds seven more transforms: AST truncation, JSON compression, log deduplication, path compression, markdown stripping, system-prompt compression, and conversation-history dedup
• AST truncation is one specific Pro transform. The roughly 78 percent figure is scoped to that single transform operating on line-numbered Python output only: it shortens function bodies while keeping their signatures and type annotations. It is not an overall or typical savings number; most content is not line-numbered Python, and the other transforms save far less
• That makes AST lossy by design, so Python code blocks are not preserved byte-for-byte; the free-tier and other Pro transforms reduce tokens without changing meaning

TokenStack Savings Tracking

The dashboard shows how many tokens the proxy is saving you, in several places.

• The proxy indicator shows a live PRO or FREE tier badge plus session and 30-day savings percentages
• The Overview header reports tokens saved for the current session and lifetime
• The Burn Report breaks compression savings down per transform and shows estimated cost, with all-time, daily, weekly, and monthly filters
• The Burn Report also shows a Per-Agent Token Cost breakdown for the Manager, Builder, and Reviewer

Model Selection

The per-agent model dropdowns on the Agent Monitor and Settings pages offer Claude Fable 5, Opus 5, and Sonnet 5 alongside the earlier Opus 4.8, Opus 4.6, Sonnet 4.6, and Haiku 4.5. Pricing for each is included in cost estimates.

4-Perspective Reviewer

The Reviewer agent evaluates every Builder change against the task specification across four dimensions before approving.

• Security, Performance, Quality, and Correctness each receive a PASS, WARN, or FAIL verdict
• Any FAIL sends the change back to the Builder with specific fix instructions
• The Reviewer reads the actual changed files rather than trusting the Builder's summary

Multi-Subtask Orchestration

When a task would touch three or more files, the Manager agent splits it into focused subtasks instead of one large change.

• Each subtask is scoped to one or two files for more reliable edits
• Subtasks run sequentially, each built and reviewed on its own
• The final report summarizes how many subtasks passed or failed