Serveur MCP Docket

Docket

One list every AI tool you use can write to — Claude Code, Codex, Cursor, Warp — across every project, before the work is worth a ticket. Local-first, self-hostable, no SaaS account.

Quick start

You need: Node.js 18+ and at least one MCP host — Claude Code, Codex, Cursor, Windsurf, Claude Desktop, Zed or Warp.

npx -y @pasichdev/docket setup

That configures every host it finds on this machine and ends by telling you which ones. Then restart your agent and ask it "add a todo: buy milk" — it shows up at http://localhost:8787, a dashboard that started by itself the moment the agent connected.

Only want Claude Code, and nothing written anywhere else? claude mcp add docket -- npx -y @pasichdev/docket does just that part. Other hosts: Supported hosts.

npm install -g @pasichdev/docket        # the hook runs a command, so it needs one on PATH
docket hook install                     # then: docket hook doctor

hook install works without the global install too — it pins the command to this exact copy of docket and tells you it did — but the short form survives moving or reinstalling, and npx leaves nothing on PATH.

Why

A thought that shows up mid-session is worth capturing but not worth the ceremony: a Notion template, a GitLab issue format, a ticket id you have to invent. So today it evaporates. Docket is the layer underneath all of that — work lands here at the speed an agent can type, from any tool, in any project, and graduates to Notion/GitLab/Obsidian when it's earned it.

Nobody else occupies this space, and it isn't an accident: Anthropic will not integrate Claude Code with Cursor, and Cursor will not integrate with Codex. Every vendor optimises its own closed loop. The space between the tools is structurally nobody's.

One list, many projects

Docket files every item under the project it was captured in, automatically — resolved from the git remote of wherever the agent is running. You never type it, and no agent has to remember to.

~/work/backend      claude-code, codex   ─┐
~/side/tracker      claude-code          ─┼──▶  docket  ──▶  one list, three scopes
~/side/notes        codex, warp          ─┘

• todo_list in ~/work/backend shows that project's open items, not all three projects' — compact, one line each.
• The web dashboard has a workspace switcher with per-project open counts.
• Items with no project context land under Unfiled and stay visible, never guessed at.

Using the git remote rather than the path means the same repo cloned to ~/src/backend on a laptop and /work/backend on a desktop is one workspace — which matters precisely because sync exists. Full resolution order and the .docket.json override: docs/workspaces.md.

Upgrading from 2.x

Read this before you upgrade if you have existing items.

3.0 migrates your store from data format v7 to v8 on first run, automatically. The migration itself is safe and is not the risk. Downgrading afterwards is.

docket 2.3.1 writes the store from its own v7 shape: it has never heard of the fields v8 adds, so its very first write after a reinstall silently strips them from every item. Nothing errors, and a later re-upgrade hands out fresh sequence numbers that no longer mean what your paired devices think they mean. 2.3.1 is published and cannot be patched, so 3.0 defends the only way it can — by keeping a copy of your pre-migration store:

~/.docket/todos.v7-pre-upgrade.enc

It is written once, before the first v8 write, and never overwritten. The upgrade prints its path on the run that creates it.

To go back to 2.x, restore it first:

docket restore --from-v7                     # puts the v7 store back; moves the v8 one aside
npm install -g @pasichdev/docket@2.3.1

In that order. restore --from-v7 deletes nothing — your v8 store is renamed aside, so you can come forward again later.

If you would rather have a portable copy as well, docket backup ./pre-v3.backup before upgrading gives you one that includes your identity and paired peers.

Bridges, not replaces

Docket is deliberately not where work lives forever. It's where work lands first, before anyone knows whether it deserves a ticket. Most of it doesn't and gets closed by hand; the rest graduates.

sourceUrl is the bridge in both directions. Set it whenever an item maps to something with a URL — a GitLab issue, a Notion page, an Obsidian share link, a Slack thread, a GitHub PR — and the card carries a clickable link straight back to it.

Why not just use GitHub Issues? Because an issue costs a title you have to phrase for an audience, a repo you have to pick, and labels you have to maintain — and because your Cursor session can't write one for you while you're mid-thought in a different project. Docket costs one sentence, from whichever tool you already have open. When the item turns out to matter, it becomes an issue, and sourceUrl remembers where it went.

Supported hosts

Any MCP host works — the tools are identical everywhere:

{
  "mcpServers": {
    "docket": { "command": "npx", "args": ["-y", "@pasichdev/docket"] }
  }
}

Host	Config file
Claude Code	claude mcp add docket -- npx -y @pasichdev/docket
Claude Desktop	~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows)
Cursor	.cursor/mcp.json or Global MCP settings
Windsurf	~/.codeium/windsurf/mcp_config.json
Codex	~/.codex/config.toml (mcp_servers)
Zed	~/.config/zed/settings.json — uses context_servers, see below

{
  "context_servers": {
    "docket": { "command": { "env": {}, "path": "npx", "args": ["-y", "@pasichdev/docket"] } }
  }
}

The one hook

Claude Code additionally gets a SessionStart hook, and only that one:

docket hook install      # writes .claude/settings.json (--global for user-wide)
docket hook doctor       # proves it fires, and reports measured latency
docket hook uninstall    # removes only the entries docket owns

When a session starts it injects the items open in that project — compact, at most 7, under 120 tokens. Nothing at all when the project has nothing open. Its whole job is continuity: you come back to a terminal and the thread is already there.

Don't want it? export DOCKET_HOOKS=off disables it immediately, without editing any config or uninstalling anything. docket hook doctor measures the real round trip and says so if it is slow enough to notice.

It fails open, always. Server not running, request timed out, malformed response, DOCKET_HOOKS=off — every one of those exits 0 and prints nothing. A tool that degrades your session when the tool itself is broken gets uninstalled, at which point it helps nobody. docket hook doctor runs the configured command for real and reports what a session would actually see, so a hook that never fires is visible rather than merely silent.

No PreToolUse, no blocking, no other hosts — the hook's only job in this release is continuity, not enforcement.

Guidance for non-Claude-Code agents

The tools work everywhere; the guidance ships as a Claude Code plugin. For any other agent, copy skills/docket/SKILL.md (everything below the --- frontmatter) into whichever file your agent reads — AGENTS.md for Codex, .cursor/rules/docket.mdc for Cursor, .windsurfrules for Windsurf, CLAUDE.md for Claude Desktop, or Warp's custom-instructions setting.

/plugin marketplace add pasichDev/docket
/plugin install docket@docket

Tools

Tool	Description
todo_add(title, description?, list?, category?, priority?, dueDate?, sourceUrl?, workspace?)	Capture an item. Filed under the current project automatically.
todo_list(filter?, list?, category?, agent?, session?, inProgress?, workspace?, verbose?, limit?, offset?)	Scoped to the current project and compact by default. workspace:"*" for everything.
todo_edit(id, ...)	Edit any subset of fields by id. Pass "" to clear an optional field.
todo_claim(id) / todo_release(id)	Mark an item in progress, or drop the claim. Auto-expires after 15 minutes.
todo_complete(id)	Mark done (also clears any claim).
todo_history(id)	Full change log for one item.
todo_delete(id)	Permanently remove an item.
todo_version() / todo_check_update()	Data-format version; read-only npm version check.
digest_publish(title, summary, sections?, metrics?, highlights?, sources?, windowFrom?, windowTo?)	Save a digest an agent compiled from your GitLab/GitHub/Notion/git — see Digests.
digest_list(limit?) / digest_get(id) / digest_delete(id)	Recent digests, one in full, remove one (everywhere it synced).
digest_take(item)	Hand item 7 (or D-7K2F9A/7) to the calling agent: its brief, plus a docket task claimed in its name.
digest_seen()	Items the user marked seen, so the next digest leaves them out.

Full field and workflow reference: skills/docket/SKILL.md.

Digests

Ask your agent "make a digest" (or "зроби дайджест"). The docket:digest skill reads your merge requests, pull requests, Notion tickets, local commits and docket items, checks every status at the source, and publishes the result — which becomes the dashboard's home page: what needs you, what shipped, what is in review, what is stuck, each item linked back to where it lives and one click away from becoming a task.

Docket never holds a GitLab, GitHub or Notion credential: the agent reads them with the CLIs and MCP servers it already has, and Docket only keeps what it wrote — on this machine in Local Mode (synced to paired devices), or on the Docket Server in Self-hosted Mode, shared by every client. Anything else the agent can reach — Jira, Linear or Sentry through their MCP servers, an Obsidian vault, a project's docs folder — can be added as a source, and the digest can be split into groups such as work, learning and side projects. What to read is local to each machine, in ~/.config/docket/digest.json (the docket:digest-setup skill writes it); the digests themselves sync to paired devices. Details: docs/digests.md.

• What changed since the previous digest is computed when it is published — new items, status changes, what dropped out — not left to the agent's memory.
• Who does what: every item can name its owner (you, a teammate, the agent); By person turns the digest into numbered next steps per person.
• Depth where it matters: routine items stay one line; blocked, failing or stale ones carry the agent's analysis of the actual cause and next step.
• Hand it off by number: tell any agent "take 7" — digest_take gives it the brief and a task claimed in its name; it closes the task with the reason.
• Seen hides an item until its status changes; marks sync across devices.
• Yours to shape: groups, people, presets ("digest work", "digest week"), environment checks, mail and chat via MCP, a daily run — and the skill keeps short notes on what you keep correcting.

CLI

docket list [-w <project>] [--all]          What's open, scoped like the MCP default
docket workspaces                           Projects, with open counts and last activity
docket sessions                             Agent sessions open right now, and where
docket stats | export | import              Inspection and round-tripping
docket hook install | uninstall | doctor    Claude Code SessionStart hook
docket serve | pair | devices | status      Self-hosted server & devices
docket backend use <url> | localize         Switch deployment mode
docket backup <file> | restore <file>       Encrypted full-device backup
docket web                                  Ensure the Web UI is running
docket check-update | update                Version management

docket help prints the canonical, always-current list. Full reference: docs/cli.md.

Deployment modes

	Local Mode	Self-hosted Mode
Default?	Yes — zero config	Opt-in
Where state lives	This machine (~/.docket)	The Docket Server you run
Setup	docket setup	docket setup, choose "Self-hosted", or docket pair <url>
Web UI	Runs on this machine	Served by the Docket Server
Multi-machine	Optional P2P sync between your own devices	Every paired device talks to one server
Claims	Advisory — a claim can be taken over, and across P2P it can be up to one pull interval stale	Atomic — a racing claim gets a 409, decided by one authority
Good for	A single machine, or a few you personally use	An always-on Raspberry Pi, mini PC, NAS, home server, or VPS
If the connection drops	N/A	Every read/write/claim fails clearly — never a silent local fallback

Both modes install from the same package and expose the same MCP tools. Full self-hosted setup and what it deliberately doesn't do: docs/self-hosting.md.

P2P sync is deprecated as of v3.0. It still works, and nothing is removed in this release — but its 15-second pull interval means claims are advisory across it, so it cannot deliver an atomic guarantee. Self-hosted Mode is the supported multi-machine path. See docs/p2p-sync.md.

Web UI

A real-time read/write dashboard — http://localhost:8787 by default in Local Mode (override with DOCKET_WEB_PORT), or the Docket Server's own URL in Self-hosted Mode. The home page (/) shows the latest digest next to the task list at a glance; the list itself is at /tasks. Workspace switcher with per-project open counts, an active- sessions panel, light/dark theme, search, sort, inline edit, undo-delete, responsive mobile layout.

In Local Mode it starts itself: the first MCP client to connect spawns it detached in the background if nothing's listening yet. Updates push live over Server-Sent Events (/api/events), no polling.

Opening it from another device on your LAN (phone, tablet) requires an explicit Viewer Gate approval from the host machine first — see Security.

Security

Docket has four separate threat models (encrypted local storage, P2P sync, the LAN Viewer Gate, and self-hosted client/server traffic) — a guarantee from one does not apply to another. The basics:

• At rest: todos.json.enc and history.json.enc are AES-256-GCM encrypted with a locally generated key — protects against accidental exposure, not against someone with read access to your own user account.
• P2P sync: X25519 ECDH + HKDF-derived per-pair secrets, HMAC-signed requests with replay protection, AES-256-GCM encrypted responses.
• LAN Viewer Gate: any browser other than the host machine needs explicit human approval before it can open the dashboard; that local traffic itself is plain HTTP, not TLS (documented tradeoff, not an oversight).
• Self-hosted mode: per-device HMAC signatures with timestamp+nonce replay protection. The server is not end-to-end encrypted against its own operator.
• Updates: every release is published with Sigstore-backed npm provenance, verifiable with npm audit signatures.

Full threat model, exact primitives and limitations: docs/security.md.

Data & encryption

Authoritative data lives on this machine (~/.docket) in Local Mode, or on the Docket Server in Self-hosted Mode — either way on infrastructure you control, never a hosted account.

• todos.json.enc — the store, AES-256-GCM encrypted
• history.json.enc — the full audit log, kept off the store's write path
• digests.json.enc — digests, AES-256-GCM encrypted, with their own sync cursor
• key — a locally generated 256-bit key, chmod 600
• device.json — this machine's id, name, and X25519 identity keypair
• peers.json.enc — paired P2P devices and their derived sync secrets
• sessions.json — which agent sessions are open right now (plain, local-only, never synced; it holds process metadata, not content)

Set DOCKET_DATA_DIR to relocate the directory; startup refuses to silently split an existing store rather than guessing.

Backup & updating

docket backup <file> bundles the whole data directory into one password-protected file (AES-256-GCM, scrypt-derived key); docket restore writes it back, renaming what's on disk aside rather than overwriting.

docket check-update   # read-only — reports current vs. latest
docket update         # checks, confirms, installs, self-tests, rolls back on failure

Testing

npm test

Runs the full node:test suite — sync delivery and pagination, the cross-process file lock under real contention, workspace resolution and scoping, the session registry, agent-facing token budgets, and the hook's fail-open behaviour.

License

MIT — see LICENSE.