Skip to main content
Glama

Narla MCP connector

Work with your Narla AI agents from the assistant you already have open: ChatGPT, Claude or Cursor.

Narla installs AI agents inside a company: Narla Marketing (edits photos and video, writes and publishes posts, keeps the website and online shop up to date), Narla Ops (company systems and documents) and Narla Voice (answers customers by phone and WhatsApp). The team directs them in plain language. This connector, built on the open Model Context Protocol, lets the team do that from any MCP-capable assistant.

What it does

Tool

Scope

What it does

list_agents

read

The agents installed for the company and what each one does

list_tasks

read

Recent tasks and their status, newest first

get_task

read

One task: its instructions, its status and the agent's report (what it did and what it changed)

read_business_memory

read

The company's single source of truth: brand kit, services, prices, hours

create_task

write

Gives an agent a task in plain language. Saved as Draft: nothing runs yet

cancel_task

write

Voids a Draft task

approve_task

approve

Approves a Draft task so the agent starts it. Requires confirm: true

Nothing runs without a person's approval. Every task starts as a Draft, and every finished task keeps a report. Narla Voice takes no instructions from staff; changes to it go through Narla Ops.

Tasks live in a Notion database (the queue the agents read). The connector only touches pages inside that one queue.

Related MCP server: Forja MCP Server

Quick start

npm install
npm run build
NARLA_DEMO=1 NARLA_TOKENS="$(openssl rand -hex 32):read,write" npm start   # in-memory demo, no Notion

With Notion, copy .env.example, fill in a Notion integration token that has access only to the task queue (and, optionally, the business memory page), and start the server.

Connect your assistant

Cursor (local, no server needed)

Clone this repository, run npm install && npm run build, then add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "narla": {
      "command": "node",
      "args": ["/path/to/narla-mcp/dist/stdio.js"],
      "env": {
        "NOTION_TOKEN": "secret_xxx",
        "NARLA_QUEUE_DATA_SOURCE_ID": "your-data-source-id",
        "NARLA_TASK_PREFIX": "NARLA-TASK",
        "NARLA_SCOPES": "read,write"
      }
    }
  }
}

Cursor can also use the remote server: "url": "https://your-host/mcp" with header Authorization: Bearer <token>.

ChatGPT and Claude (remote)

Deploy the HTTP server (any Node 20+ host). In the assistant's custom connector settings, add:

https://your-host/mcp/<your-token>

The token in the path is for apps whose connector settings take only a URL. Treat that URL as a password. Clients that support headers should use Authorization: Bearer <token> with https://your-host/mcp instead.

Claude Desktop (local)

Same as Cursor, in claude_desktop_config.json.

Configuration

Variable

Required

Meaning

NOTION_TOKEN

yes, unless NARLA_DEMO=1

Notion integration token

NARLA_QUEUE_DATA_SOURCE_ID

yes

Data source id of the task queue

NARLA_TASK_PREFIX

no

Task numbering prefix, e.g. NARLA-TASK gives NARLA-TASK-12: ...

NARLA_MEMORY_PAGE_ID

no

Notion page with the business memory

NARLA_DEFAULT_MODEL

no

Value for the queue's Model select on new tasks

NARLA_TOKENS

HTTP only

token:scopes pairs separated by ;. Tokens must be 32+ characters

NARLA_SCOPES

stdio only

Scopes for local mode. Default read,write

HOST, PORT, ALLOWED_HOSTS

HTTP only

Bind address and allowed Host headers

The queue is expected to have these properties: Title (title), Status (select with Draft, Approved, Running, Done, Failed, Voided), Tier (select: Low, Medium, High), Result (text, written by the agent), and optionally Model (select).

Security

  • Give each person their own token, with the smallest scope they need. Only the owner should have approve.

  • Tokens are held in memory only as SHA-256 hashes and compared in constant time.

  • The Notion integration should be shared with the queue and the memory page only. The connector refuses any page outside the configured queue.

  • The HTTP server is stateless and keeps no conversation data.

  • OAuth is on the roadmap; until then, rotate a token by changing NARLA_TOKENS and restarting.

Found a security issue? Write to contact@narla.digital rather than opening a public issue.

Development

npm test        # vitest: tools, scopes, HTTP auth, Notion store
npm run dev     # HTTP server with tsx

License

Apache License 2.0. Copyright 2026 NARLA S.R.L.

Available Tools

6 tools
cancel_taskCancel a task that has not startedA
DestructiveIdempotent

Marks a Draft task as Voided, so it never runs.

Inputs: task_id (required).

create_taskGive an agent a taskA

Creates a task for one agent, in plain language. The task is saved as a Draft and does not run until a person approves it. Narla Voice does not take tasks; send changes for it to Narla Ops.

Inputs: agent (required), title (required), instructions (required), tier.

get_taskRead a task and its reportA
Read-only

Returns one task: its instructions, its status and the report the agent wrote (what it did and what it changed).

Inputs: task_id (required).

list_agentsList the company's agentsA
Read-only

Lists the Narla agents installed for this company and what each one does.

Inputs: no named parameters.

list_tasksList tasksA
Read-only

Lists the most recent tasks, newest first, with their status. Filter by status to see, for example, what is waiting for approval (Draft) or what is finished (Done).

Inputs: limit, status.

read_business_memoryRead the business memoryA
Read-only

Returns the company's business memory, the single source of truth every agent reads: brand kit, services, prices, opening hours and approved answers.

Inputs: no named parameters.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool targets a distinct resource and action: list_agents (agents), list_tasks/get_task (tasks at collection vs single level), read_business_memory (memory), and create_task/cancel_task (distinct mutations). There is no overlap that would cause misselection.

    Naming Consistency5/5

    All six tools follow a clean verb_noun pattern (list_agents, list_tasks, get_task, read_business_memory, create_task, cancel_task) in consistent snake_case. Minor get/read variation is idiomatic and predictable.

    Tool Count5/5

    Six tools is well-scoped for an agent/task connector: enough to cover discovery, inspection, creation, and cancellation without bloat.

    Completeness3/5

    The task surface has create, list, get, and cancel, but no update tool and no approve tool, even though the workflow explicitly depends on a person approving Drafts before they run. Agents hit a dead end after creating a task, unable to approve or edit it.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Bring BasicOps into your AI assistant. With MCP, your assistant can read your workspace, turn chat into structured tasks, and keep projects moving by updating owners, due dates, and status as decisions happen. BasicOps keeps humans and agents aligned in one place—tasks, messages, and projects—so work doesn’t get lost across tools. All actions respect your BasicOps access controls.
      1
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Exposes specialized agents as MCP tools so users can manage and triage Jira tasks, analyze repositories for health and conventions, generate commit/PR/DevOps patterns, assemble reports, produce slides and infographics, and run semantic search over a knowledge base. Everything is invokable directly from inside a coding session in Claude Code, Cursor, or Claude Desktop.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents such as ChatGPT, Claude, and local LLMs to operate an enterprise digital-operations platform through MCP, exposing 19 model tools and 3 app-only tools for identity, data, workflow, files, and PII-controlled operations with role-based permissions and two-phase confirmations.
      Apache 2.0