Best for
- Developers who need task continuity between AI coding agent sessions
- Teams that prefer local, reviewable files over cloud services
- Users of Claude Code, Cursor, or Codex seeking consistency across sessions
A
Agentpack
Community
Local-first task-state ledger for AI coding agents
- Category
- Dev Tools #20 of 438
- Stars
- ★ 18 Popular
- Transport
- stdio (local process)
- Runtime
- Node.js 20+
- Credentials
- No credential needed
- License
- MIT
- Last commit
68FMRS · C
Agentpack is a specialized MCP server for privacy-focused developers who value task continuity among various local AI agent sessions. It offers strong concepts like Task Passports and checkpoints but lacks a fully documented tool list in official sources, making undocumented initial integration a challenge.
Strongest · Security and permissions 16/20 Weakest · Reliability 11/20
Why each score
- Reliability 11/20
- README claims CI badge, npm test, MCP smoke test, and a documented tool contract; zero runtime dependencies reduces failure surface. However, only README/manifest were reviewed — no test files or CI workflow content verifiable — so under static calibration it cannot exceed 12.
- Security and permissions 16/20
- Strong stated posture: local-first files, no network/telemetry, npm provenance with Trusted Publisher, ignore-scripts, committed lockfile, secret redaction; task gate warns by default rather than blocking. No red-line issues found. Deducted because these are claims not verifiable from source, and side effects of hooks writing client configs are not fully itemized.
- Maintenance 13/20
- v1.6.0, MIT license, CI badge, SECURITY.md, documented release process, and provenance indicate active maintenance. Low stars but not abandoned; no verifiable issue-response history, so no higher score.
- Documentation 15/20
- Well-layered docs (INTEGRATIONS, CLI, MCP, TASK-PASSPORT, DEMOS, VISION, SECURITY) with examples and disclosed limits, but actual tool parameter references and troubleshooting details not directly evidenced.
- Setup experience 13/20
- Clear path: global npm install, per-repo init, per-client install command (codex|claude|cursor|claude-desktop), doctor for verification, Node 20+. Deducted since concrete client config examples are not shown in the supplied material and reconnection is manual.
Static review · not runListed 2026-09-08
Read the FMRS scoring method →Fit and risk
What it can accessReads local filesWrites / deletes local files
Not for
- Teams needing cloud sync or collaboration features (it is local-first)
- Users not on Node.js 20+ or who prefer a single binary
- Those who want integration with issue trackers or project management tools (not provided)
Required permissions
- Access to the local repository filesystem (reads/writes .agentpack/)
- Global npm package installation (agentpack-cli)
- May configure git hooks or client hooks for task gating
Risks and side effects
- Mistakes in task state could lead to incorrect assumptions by agents
- Local files could be accidentally modified or lost
- Secret redaction is best-effort and not guaranteed
- Task gate could block legitimate edits if over-restrictive
Setup
Before you start
Runtime：Node.js 20+
Requires Node.js >= 20. Install agentpack-cli globally via npm. In your repository, run agentpack init. Install client config with agentpack install claude --write (or codex or cursor). Restart/reconnect the coding-agent client, then the agent loads context at session start. Verify with agentpack doctor.
Check that it works
Run agentpack init and agentpack install claude --write in your repo, restart the client, then run agentpack doctor to verify the setup and confirm the agent loads Agentpack context at session start.
Troubleshooting
- Run agentpack doctor to verify setup.
- Ensure Node.js >= 20.
- Restart/reconnect the client after installation.
- Use CLI commands like agentpack resume --preset agent --query "topic" to inspect state. Check docs/INTEGRATIONS.md for client-specific settings.
Things to try
Once connected, you can ask your AI assistant things like:
- Load the Agentpack task state and continue where the last session left off
- Record the architecture decision we just made into the task ledger
- Create a checkpoint in .agentpack/ with current status and next actions
- Check whether my current edits bypass the active task scope
Use cases
Working on a long-running refactor or bugfix across sessions
Switching between Claude Code, Cursor, Codex, or multiple MCP clients
Returning to a task days later without context
Keeping agents focused on a scoped task in a monorepo with task gates
Overview
Agentpack is a local-first MCP server that preserves task continuity for AI coding agents. It maintains a reviewable task ledger (.agentpack/) inside your repository, recording goals, decisions, dead ends, verification evidence, and checkpoints. Agents record durable state while working through MCP tools, and any next session — the same agent after compaction, a different client, or a future you — loads that state and continues from where it left off. Agentpack also generates project instructions (AGENTS.md, CLAUDE.md, Cursor rules) to guide when agents should load and write state. A CLI provides the same state for human inspection and manual handoff. Core features include task passports, task gates (hooks that warn or block edits outside the active task), and budgeted context output.
Similar servers
Context7
80 · B
Upstash's official server providing up-to-date third-party library docs for AI coding assistants
NPM Sentinel MCP Server
79 · B
AI-powered NPM package analysis MCP server
MartinLoop MCP Server
79 · B
An independent control layer that gives AI coding agents budgets, brakes, and receipts
GitHub MCP Server
76 · B
GitHub's official MCP server for managing repos, issues, PRs, and workflows via natural language
Source revision ee35e4aee60e Data synced 2026-10-19 Read the FMRS scoring method