Tack

tack

[](https://www.npmjs.com/package/tack-cli) [](LICENSE)

Tack gives your agents a shared memory for the repo, so each new session can pick up where the last one left off.

Static instruction files go out of date as soon as the code changes. Tack keeps a shared record in ./.tack/ and checks that memory against the codebase, so the next agent starts with context that still matches reality.

Install

1. Install Tack. If you prefer not to install globally, use npx tack-cli@latest instead of tack.

npm install -g tack-cli

1. Initialize the repo and install your agent startup instructions once.

tack init
tack setup-agent

1. Keep tack watch open in a second terminal so you can see what the agent is doing.

tack watch

1. Start the MCP server from the repo root with a visible agent label.

TACK_AGENT_NAME=claude tack mcp

On Windows PowerShell, set the label like this instead:

$env:TACK_AGENT_NAME="claude"
tack.cmd mcp

1. Check tack watch for these events:

• READY means the MCP session connected
• READ means the agent read tack://session or another Tack context resource
• WRITE means the agent wrote memory back with checkpoint_work, log_decision, or log_agent_note

If you see READY, READ, and WRITE, the agent is actually using Tack.

When you wrap up a session, create a handoff:

tack handoff

Read the full guides:

• [Product Direction](./docs/product-direction.md)
• [Getting Started](./docs/getting-started.md)
• [Agent Workflow](./docs/agent-workflow.md)
• [MCP Clients](./docs/mcp-clients.md)
• [Watch Mode](./docs/watch.md)
• [CLI Reference](./docs/cli.md)
• [Detectors And YAML Rules](./docs/detectors.md)

What Tack Actually Does

• Keep implementation facts aligned with the real codebase
• Stop stale project instructions from misleading the next agent
• Preserve decisions, blockers, discoveries, and partial work across sessions
• Show when agents are reading current context and writing memory back
• Carry context across sessions with handoffs and MCP context resources

Why This Matters

• The bigger problem is stale context, not just missing context
• Every repeated question like "what framework are we using?" is a context failure
• Drift detection helps keep Tack honest when the code has changed
• Decisions explain why the system looks the way it does before the agent asks
• Handoffs let context survive session boundaries without another interview

Typical Workflow

Initialize Tack once at the repo root:

tack init

Install agent instructions once per repo:

tack setup-agent

When you connect an agent, run:

tack watch
TACK_AGENT_NAME=claude tack mcp

On Windows PowerShell:

tack watch
$env:TACK_AGENT_NAME="claude"
tack.cmd mcp

tack watch shows whether the agent connected, read context, and wrote work back. The usual flow is READY, then READ, then WRITE.

Example:

[READY][claude] connected to Tack MCP
[READ][claude] read session context
[WRITE][claude] checkpointed work

There is no standalone tack check-in command in v1. Write-back happens through MCP tools like checkpoint_work, plus tack handoff.

During normal work:

tack status
tack watch

If you use more than one agent, give each MCP server its own TACK_AGENT_NAME so tack watch can show who read context, who checked rules, and who wrote memory back.

Example Workflows

New Agent Session

Open a new agent window in a repo that already uses Tack:

tack watch
TACK_AGENT_NAME=claude tack mcp

On Windows PowerShell:

tack watch
$env:TACK_AGENT_NAME="claude"
tack.cmd mcp

The agent reads tack://session, sees the current focus and recent work, and starts with current context instead of re-learning the repo. tack watch shows that read as it happens, so you can tell the agent is working from real project memory.

Structural Change

The agent wants to add a dependency or introduce a new boundary.

Instead of guessing from old instructions, it calls check_rule first. That gives a quick guardrail check before the architecture changes.

End Of Session

The agent made changes, found one blocker, and left partial work.

Instead of making the next session piece things together, it calls checkpoint_work once. Tack saves a summary, discoveries, decisions, and related files so the next session can pick up quickly.

Learn More

• [Product Direction](./docs/product-direction.md)
• [Getting Started](./docs/getting-started.md)
• [Agent Workflow](./docs/agent-workflow.md)
• [MCP Clients](./docs/mcp-clients.md)
• [Watch Mode](./docs/watch.md)
• [CLI Reference](./docs/cli.md)
• [Detectors And YAML Rules](./docs/detectors.md)

Notes

• offline only, no hidden network calls in Tack's project-memory engine
• writes are guarded to ./.tack/
• Python virtual environments are ignored during scans (venv, .venv, site-packages) to reduce false positives

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: khuper
• Source: khuper/tack
• License: MIT

Install and usage instructions live in the source repository linked above.