guanling
A task-management MCP server that enforces strict gates on task states, WIP limits, and closing reasons, refusing operations that would leave the list ambiguous or rotting.
README
guanling (關令)
A task-management MCP server that refuses.
Most task tools do what you tell them. This one has opinions and enforces them.
Why this exists
A real task list grew to 280 items. Sixteen were marked "in progress". Not one of them was actually being worked on.
The missing thing was not a tool — there already was a list. The missing thing was anything that forced a decision. Every operation said yes, so the list accumulated intentions and never spent them.
What makes Linear work is not its feature set. It is its refusals: a small fixed set of states you cannot extend, and limits you cannot argue with. This server copies the refusals, not the features.
The gates
Each one is a refusal, not a warning.
| Gate | Behaviour |
|---|---|
owner and next_step required |
task_add is rejected without both. No owner and no next action means it is a wish, not a task. |
| Five states, hardcoded | triage · todo · doing · done · dropped. There is no API to add a sixth. Custom statuses are how a list starts rotting. |
WIP limit on doing |
Moving past the limit is rejected, and the error names what currently occupies it so you have to close something first. |
| External items land in triage | Anything with a source starts in triage, never straight in todo. |
| Closing requires a reason | done / dropped without reason is rejected. Why it closed outlives that it closed. |
| Rollover is visible | cycle_roll increments a rolled counter and returns anything rolled 3+ times as needing a decision. Nothing is dropped silently. |
Refused calls write nothing to disk — there is a test for this.
Install
npm install -g guanling
Then add it to your MCP client config:
{
"mcpServers": {
"guanling": {
"command": "npx",
"args": ["-y", "guanling"],
"env": {
"GUANLING_AGENT": "me",
"GUANLING_FILE": "/path/to/my-tasks.json",
"GUANLING_WIP": "3"
}
}
}
}
| Variable | Default | Meaning |
|---|---|---|
GUANLING_AGENT |
unknown |
Identity for this instance. |
GUANLING_FILE |
~/.guanling/<agent>.json |
Path to the JSON store. |
GUANLING_WIP |
3 |
Maximum concurrent doing items. |
⚠️ Give every instance its own GUANLING_FILE. Two processes sharing one
file will overwrite each other — see Limitations.
Tools
| Tool | Purpose |
|---|---|
task_add |
Open a task. Rejects without owner + next_step. |
task_update |
Change status/owner/next_step. Enforces the WIP limit and the close-reason rule. |
task_list |
List open work. stale_days surfaces tasks opened and never touched. |
task_triage |
Resolve inbox items: accept / decline / duplicate / snooze. |
cycle_status |
Days left, WIP usage, stale count, and what has rolled twice or more. |
cycle_roll |
Close the cycle, carry unfinished work forward with a visible counter. |
Storage
One JSON file, written atomically (write to a temp file, then rename).
Deliberately not a database. The scale is hundreds of rows, and a plain file
stays readable by a human, diffable in git, and backed up by cp. The complete
string is built before the file is opened, so an exception mid-serialisation
leaves the previous file intact.
Tests
node selftest.mjs
16 arms, weighted toward the refusals — those are the product. Includes a
negative control asserting that a rejected task_add leaves nothing on disk.
The WIP gate is mutation-tested: removing it turns exactly the two arms that cover it red. A gate that cannot be observed failing has not been verified.
Limitations
Stated up front rather than discovered later.
- No concurrency lock. One process per store file. Two writers will clobber each other.
- No cross-instance visibility. Each instance sees only its own file. This is intentional for per-agent isolation, but it means separate agents cannot see each other's work and may duplicate effort.
snoozesetssnooze_untilbut nothing wakes it. Items do not return on their own;task_listis how you find them.- No history or audit trail. The file holds current state only.
The name
關令尹喜 was the gatekeeper at Hangu Pass. When Laozi tried to leave through it, the gatekeeper would not let him pass until he wrote his teaching down — which is why the Tao Te Ching exists.
Same idea: you do not get through until you write it down.
Privacy Policy
What is collected: nothing. Guanling has no telemetry, no analytics, no crash reporting, and no update check.
What is stored, and where: only the tasks you create — their title, owner,
next step, status, optional source and blocker, timestamps, and close reason.
They are written to a single JSON file on your own machine, at the path you
configure (GUANLING_FILE, or ~/.guanling/<agent>.json by default). Nothing
is written anywhere else.
Third-party sharing: none by guanling itself. The server makes no network
requests: no HTTP client, no sockets, no telemetry. It uses four Node builtins —
node:fs, node:path, node:crypto (to mint delegation nonces), and
node:child_process.
⚠️ That last one matters and is stated plainly: if you configure
GUANLING_SEND_CMD, guanling executes your command and pipes a delegation
payload to it on stdin. Whatever that command does with the payload — including
sending it over a network — is outside guanling and is your choice. With the
variable unset, delegation is refused and nothing is ever executed.
(Earlier releases of this file claimed two builtins. That stopped being true in 1.1.0 when delegation was added, and the sentence was not updated until 1.4.0.)
Data retention: entirely yours. The file persists until you delete it.
Guanling never expires, prunes, or deletes data on its own — closing a task
marks it done or dropped and keeps the record. To erase everything, delete
the JSON file.
Your control: the store is plain JSON. You can read it, edit it, copy it, back it up, or delete it with ordinary file tools. No export feature is needed because the file is the export.
Contact: open an issue at https://github.com/fateLiang/guanling/issues.
License
MIT
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.