@selfagency/beans-mcp 🫘

NPM Version

MCP (Model Context Protocol) server for Beans issue tracker. Provides programmatic and CLI interfaces for AI-powered interactions with Beans workspaces.

Documentation: beans-mcp.self.agency

🤖 Try Beans fully-integrated with GitHub Copilot in VS Code! Install the selfagency.beans-vscode extension.

Usage

Terminal

npx @selfagency/beans-mcp /path/to/workspace

Versioning

@selfagency/beans-mcp has its own package versioning. Compatibility with the Beans CLI is tracked separately.

At startup, the server compares the installed beans CLI version against the hardcoded supported Beans version: 0.4.2. If they differ, it prints a warning to stderr and continues startup.

Parameters

--workspace-root or positional arg: Workspace root path
--cli-path: Path to Beans CLI
--port: MCP server port (default: 39173)
--log-dir: Log directory
-h, --help: Print usage and exit

Summary of public MCP tools

Tool	Description
`beans_init`	Initialize the workspace (optional `prefix`).
`beans_archive`	Archive completed/scrapped beans.
`beans_view`	Fetch full bean details by `beanId` or `beanIds`.
`beans_create`	Create a new bean (title/type + optional body/parent).
`beans_bulk_create`	Create multiple beans in one call, optionally under a shared parent.
`beans_update`	Consolidated metadata + body updates (status/type/priority/parent/clearParent/blocking/blockedBy/body/bodyAppend/bodyReplace) plus optional optimistic concurrency hint (`ifMatch`).
`beans_bulk_update`	Update multiple beans in one call, optionally reassigning them to a shared parent.
`beans_complete_tasks`	Mark all markdown checklist tasks within a bean as complete.
`beans_delete`	Delete one or many beans (`beanId` or `beanIds`, optional `force`).
`beans_reopen`	Reopen a completed or scrapped bean to an active status.
`beans_query`	Unified list/search/filter/sort/ready operations, with GraphQL passthrough.
`beans_bean_file`	Read/edit/create/delete files under `.beans`.
`beans_output`	Read extension output logs or show guidance.

Notes

The beans_query tool is intentionally broad: prefer it for listing, searching, filtering or sorting beans, and for generating Copilot instructions (operation: 'llm_context').
All file and log operations validate paths to keep them within the workspace or the VS Code log directory. The .beans/ prefix is automatically stripped from paths — you can pass either some-bean.md or .beans/some-bean.md and the result is the same.
beans_update replaces many fine-grained update tools; callers should use it to keep the public tool surface small and predictable.
beans_archive provides CLI parity for archiving completed/scrapped beans.
Closing a parent bean via beans_update (status: completed or status: scrapped) cascades the same status to all descendants.
Reopening a parent bean via beans_reopen cascades the target status to closed descendants (completed / scrapped).
beans_bulk_create and beans_bulk_update are best-effort: they process each item sequentially and return a per-item result array with success/error entries rather than failing atomically.
Frontmatter title: values are automatically double-quoted on write. Pass raw titles — quoting and escaping is handled for you.
beans_bean_file supports update_frontmatter for atomic frontmatter-only writes; supported fields include pr and branch.
Unfiltered list results are cached with a short burst TTL and a timestamp-probe refresh strategy. Mutation tools (beans_create, beans_update, beans_delete, etc.) invalidate the cache immediately.
Version mismatches between beans-mcp and the Beans CLI are warning-only and non-blocking by design.
When beanId is missing in tool input, validation errors include a hint: Did you mean \beanId`?`.

Examples

beans_init

Request:

config.json

{ "prefix": "project" }

Response (structuredContent):

config.json

{ "initialized": true }

beans_view

Request:

config.json

{ "beanId": "bean-abc" }

Request (multiple beans):

config.json

{ "beanIds": ["bean-abc", "bean-def"] }

Response (structuredContent):

config.json

{
  "bean": {
    "id": "bean-abc",
    "title": "Fix login timeout",
    "status": "todo",
    "type": "bug",
    "priority": "critical",
    "body": "...markdown...",
    "createdAt": "2025-12-01T12:00:00Z",
    "updatedAt": "2025-12-02T08:00:00Z"
  }
}

beans_archive

Request:

config.json

{}

Response (example):

config.json

{ "archived": true, "archivedCount": 3 }

beans_create

Request:

config.json

{
  "title": "Add dark mode",
  "type": "feature",
  "status": "todo",
  "priority": "normal",
  "body": "Implement theme toggle and styles",
  "parent": "epic-123"
}

description is accepted as a deprecated alias for body.

Response (structuredContent):

config.json

{
  "bean": {
    "id": "new-1",
    "title": "Add dark mode",
    "status": "todo",
    "type": "feature"
  }
}

beans_bulk_create

Request:

config.json

{
  "parent": "epic-123",
  "beans": [
    { "title": "Design mockups", "type": "task" },
    { "title": "Implement API", "type": "task", "priority": "high" },
    { "title": "Write tests", "type": "task", "parent": "epic-456" }
  ]
}

The top-level parent is applied as a default to any bean that does not specify its own parent. Here Design mockups and Implement API are assigned to epic-123; Write tests overrides with epic-456.

Response (structuredContent):

config.json

{
  "requestedCount": 3,
  "successCount": 3,
  "failedCount": 0,
  "results": [
    { "bean": { "id": "task-1", "title": "Design mockups" } },
    { "bean": { "id": "task-2", "title": "Implement API" } },
    { "bean": { "id": "task-3", "title": "Write tests" } }
  ]
}

beans_bulk_update

Request (move a batch of tasks to in-progress and assign them to a parent):

config.json

{
  "parent": "epic-123",
  "beans": [
    { "beanId": "task-1", "status": "in-progress" },
    { "beanId": "task-2", "status": "in-progress" },
    { "beanId": "task-3", "status": "in-progress", "parent": "epic-456" }
  ]
}

Response (structuredContent):

config.json

{
  "requestedCount": 3,
  "successCount": 3,
  "failedCount": 0,
  "results": [
    { "beanId": "task-1", "bean": { "id": "task-1", "status": "in-progress" } },
    { "beanId": "task-2", "bean": { "id": "task-2", "status": "in-progress" } },
    { "beanId": "task-3", "bean": { "id": "task-3", "status": "in-progress" } }
  ]
}

Both bulk tools are best-effort: partial failures are reported per-item rather than aborting the whole batch.

beans_update

Request (change status and add blocking):

config.json

{
  "beanId": "bean-abc",
  "status": "in-progress",
  "blocking": ["bean-def"],
  "ifMatch": "etag-value"
}

Request (atomic body modifications):

config.json

{
  "beanId": "bean-abc",
  "bodyReplace": [
    { "old": "- [ ] Task 1", "new": "- [x] Task 1" },
    { "old": "- [ ] Task 2", "new": "- [x] Task 2" }
  ],
  "bodyAppend": "## Summary\n\nAll checklist items completed."
}

Note: body (full replacement) cannot be combined with bodyAppend or bodyReplace in the same request.

Response (structuredContent):

config.json

{
  "bean": {
    "id": "bean-abc",
    "status": "in-progress",
    "blockingIds": ["bean-def"]
  }
}

beans_delete

Request:

config.json

{ "beanId": "bean-old", "force": false }

Response:

config.json

{ "deleted": true, "beanId": "bean-old" }

Batch request:

config.json

{ "beanIds": ["bean-old", "bean-older"], "force": false }

Batch response (summary):

config.json

{
  "requestedCount": 2,
  "deletedCount": 2,
  "failedCount": 0,
  "results": [
    { "beanId": "bean-old", "deleted": true },
    { "beanId": "bean-older", "deleted": true }
  ]
}

beans_reopen

Request:

config.json

{
  "beanId": "bean-closed",
  "requiredCurrentStatus": "completed",
  "targetStatus": "todo"
}

Response:

config.json

{ "bean": { "id": "bean-closed", "status": "todo" } }

beans_complete_tasks

Request:

config.json

{ "beanId": "bean-abc" }

Response:

config.json

{
  "bean": {
    "id": "bean-abc",
    "status": "todo"
  },
  "totalTaskCount": 5,
  "updatedTaskCount": 3,
  "unchangedTaskCount": 2
}

beans_query examples

Refresh (list all beans):

config.json

{ "operation": "refresh" }

Response (partial):

config.json

{ "count": 12, "beans": [] }

Filter (statuses/types/tags):

config.json

{
  "operation": "filter",
  "statuses": ["in-progress", "todo"],
  "types": ["bug", "feature"],
  "tags": ["auth"]
}

Search (full-text):

config.json

{ "operation": "search", "search": "authentication", "includeClosed": false }

Sort (modes: status-priority-type-title, updated, created, id):

config.json

{ "operation": "sort", "mode": "updated" }

Ready (actionable beans only):

config.json

{ "operation": "ready" }

LLM context (generate Copilot instructions; optional write-to-workspace):

config.json

{ "operation": "llm_context", "writeToWorkspaceInstructions": true }

Response (structuredContent):

config.json

{
  "graphqlSchema": "...",
  "generatedInstructions": "...",
  "instructionsPath": "/workspace/.github/instructions/beans-prime.instructions.md"
}

Raw GraphQL passthrough (CLI parity with beans query):

config.json

{
  "operation": "graphql",
  "graphql": "{ beans(filter: { type: [\"bug\"] }) { id title status } }"
}

With variables:

config.json

{
  "operation": "graphql",
  "graphql": "query($q: String!) { beans(filter: { search: $q }) { id title } }",
  "variables": { "q": "authentication" }
}

beans_bean_file

Request (read):

config.json

{ "operation": "read", "path": "beans-vscode-123--title.md" }

Response:

config.json

{
  "path": "/workspace/.beans/beans-vscode-123--title.md",
  "content": "---\n...frontmatter...\n---\n# Title\n"
}

Request (atomic frontmatter update):

config.json

{
  "operation": "update_frontmatter",
  "path": "beans-vscode-123--title.md",
  "fields": {
    "status": "in-progress",
    "pr": "123",
    "branch": "feature/cascade-status-and-skills-npm"
  }
}

Response:

config.json

{
  "path": "/workspace/.beans/beans-vscode-123--title.md",
  "bytes": 256,
  "updatedFields": ["status", "pr", "branch"],
  "frontmatter": {
    "status": "in-progress",
    "pr": "123",
    "branch": "feature/cascade-status-and-skills-npm"
  }
}

beans_output

Request (read last 200 lines):

config.json

{ "operation": "read", "lines": 200 }

Response:

config.json

{
  "path": "/workspace/.vscode/logs/beans-output.log",
  "content": "...log lines...",
  "linesReturned": 200
}

Programmatic usage

Installation

Terminal

npm install beans-mcp

Example

server.ts

import { createBeansMcpServer, parseCliArgs } from '@selfagency/beans-mcp';

const server = await createBeansMcpServer({
  workspaceRoot: '/path/to/workspace',
  cliPath: 'beans', // or path to beans CLI
});

// Connect to stdio transport or your own transport

API

createBeansMcpServer(opts)

Creates and initializes a Beans MCP server instance.

Options:

workspaceRoot (string): Path to the Beans workspace
cliPath (string, optional): Path to Beans CLI executable (default: 'beans')
name (string, optional): Server name (default: 'beans-mcp-server')
version (string, optional): Server version
logDir (string, optional): Directory for server logs
backend (BackendInterface, optional): Custom backend implementation

Returns: { server: McpServer; backend: BackendInterface }

startBeansMcpServer(argv)

CLI-compatible entrypoint for launching the server.

Utility Functions

parseCliArgs(argv: string[]): Parse CLI arguments
isPathWithinRoot(root: string, target: string): boolean: Check if path is contained within root
sortBeans(beans, mode): Sort beans by specified mode

Types & Schemas

Export of GraphQL schema, Zod validation schemas, and TypeScript types for Beans records and operations.

Agent Skills (`skills-npm`, `skills.sh`)

This package ships a built-in Agent Skill under skills/ and also publishes that skill in a format that fits the broader open skills ecosystem surfaced by skills.sh.

Skill path in package: skills/beans-mcp/SKILL.md
Published skill artifact: https://beans-mcp.self.agency/.well-known/agent-skills/beans-mcp/SKILL.md
Published discovery index: https://beans-mcp.self.agency/.well-known/agent-skills/index.json
Compatible with discovery tools that scan: node_modules/**/skills/*/SKILL.md

That means you can use it with npm-based workflows such as skills-npm, while also pointing ecosystem tooling at the published skill artifact and discovery index used by skills catalogs like skills.sh.

To symlink installed npm-packaged skills into your agent workspace, you can use skills-npm in your consuming project.

License

MIT

@selfagency/beans-mcp 🫘

Usage

Versioning

Parameters

Summary of public MCP tools

Examples

Programmatic usage

Installation

Example

API

createBeansMcpServer(opts)

startBeansMcpServer(argv)

Utility Functions

Types & Schemas

Agent Skills (skills-npm, skills.sh)

License

Agent Skills (`skills-npm`, `skills.sh`)