cob-portal-mcp
OfficialThis server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
cob-portal-mcp
Open the cob-portal-mcp guide page → Install, the safety model, every tool, the key scopes and the FAQ, on one page.
npm · Releases · Changelog · Callobuzz open source
An MCP server for the Call O Buzz portal. It lets an AI agent (Claude Code, Claude Desktop, or any MCP client) read tasks and send drafts and notes to the portal. It runs on your computer and talks to the portal API with a personal API key.
The agent can only make drafts and notes. It cannot assign work, approve a draft, change a task, pay, send mail or delete anything. A person must approve every draft in the portal before it becomes a task, and that person decides who gets it. Every note it writes is marked as written through the key.
Why
A good agent can plan work, write a test plan or break a goal into tasks. The portal is where a team tracks that work. This server connects the two, and keeps a person in charge:
The agent reads. It lists tasks, templates and categories, and reads one task with its steps, answers and notes.
The agent proposes. It sends draft tasks (a main task with sub-tasks) or a test plan as a check list draft.
A person decides. The drafts wait in the portal. A person approves, rejects and assigns them.
The key sets the reach. The server offers only the tools that the scopes of the key allow.
Related MCP server: Booqable MCP Server
Prerequisites
Node.js 22 or newer.
A portal account with two-step sign-in (2FA) on.
A personal API key from the portal (see Make a key).
Install in one command
Make a key (see Make a key). Copy it.
Add the server to your AI agent. npx downloads it from npm and starts it. Nothing is built on your computer.
Claude Code
claude mcp add cob-portal \
--env COB_PORTAL_URL=https://your-portal-address \
--env COB_PORTAL_KEY=cob_your_key_here \
-- npx -y cob-portal-mcpClaude Desktop
Add this to claude_desktop_config.json, then restart Claude Desktop:
{
"mcpServers": {
"cob-portal": {
"command": "npx",
"args": ["-y", "cob-portal-mcp"],
"env": {
"COB_PORTAL_URL": "https://your-portal-address",
"COB_PORTAL_KEY": "cob_your_key_here"
}
}
}
}The portal's Profile page shows this block with your new key and the portal address filled in ("Copy MCP setup").
Check it: ask the agent to run
whoami. It answers with your name, the key name and its scopes.
Notes:
To stay on one version, add it:
cob-portal-mcp@0.9.0.To run it straight from GitHub instead of npm, use
github:callobuzz/cob-portal-mcp(or#v0.9.0for one version).To use a folder copy, clone the repository and write
node <path>/dist/cli.jsin place ofnpx -y .... On Windows write the path with forward slashes, for exampleJ:/callobuzz/mcp-servers/cob-portal-mcp/dist/cli.js.
Why it works with no build: the built files in dist/ are in the npm package and committed to the repository, and bin
points at dist/cli.js. npx installs the two runtime packages (@modelcontextprotocol/sdk, zod) and runs it. There is no
prepare script, so an install never needs the build tools.
Settings
Variable | Needed | Meaning |
| yes | The portal address, as the Profile page shows it. It must be https; plain http is allowed only for |
| yes | A personal key that starts with |
| no | How long to wait for one portal answer. Default |
Make a key
Sign in to the portal. Your account needs two-step sign-in (2FA) on.
Open Profile, then API keys.
Make a new key. Choose its scopes and how long it lasts. Copy it at once. The portal shows it only one time.
Put it in
COB_PORTAL_KEY. A key acts as you, cut down to its scopes. It never has more rights than you.
If a key stops working, make a new one on the same page.
Scopes
Scope | What it allows |
| Read tasks, task templates and task categories. |
| Send draft tasks and check lists for approval. Read your draft batches. |
| Write notes on tasks. |
| Read check list results. |
| Read the current portal set-up (settings, job roles, people and their roles, and more). Changes nothing. |
| Draft change sets for a person to approve. Read, withdraw and list your own sets. Never approves, applies or undoes. |
| Read people's records: work time, hiring, leave, attendance, roster, presence, tests, targets, letters, documents and pay (pay for the admin key only). Changes nothing. |
At start the server calls GET /api/v1/me and offers only the tools whose scope the key has. whoami needs no scope.
Tools
Tool | Scope | What it does |
| none | Shows the person, the key name, its scopes and its end date. |
|
| Lists tasks ( |
|
| Reads one task with its steps, answers, notes, category, parent and sub-task count. |
|
| Lists the published task templates, with their category. |
|
| Lists the task categories the portal offers. |
|
| Reads the test runs and skips of a check list task. File names only. |
|
| Writes a note on a task. |
|
| Sends a batch of draft tasks. A batch can hold main tasks with their sub-tasks (see below). |
|
| Sends a test plan (plan JSON) as one draft with a check list. |
|
| Lists the draft batches you sent. |
|
| Reads one draft batch, with the state, category and parent of each draft. |
|
| Lists the areas you may draft for, the kinds of change and the shape of each value. Call it first. |
|
| Reads settings ( |
|
| Reads job roles with their responsibilities. |
|
| Reads people: main role, extra roles, responsibilities, designation, manager ( |
|
| Reads hiring rounds. |
|
| Reads leave types. |
|
| Reads offer letters ( |
|
| Reads the offer letter template: sections, fields, defaults. Admin key only. |
|
| Reads holidays ( |
|
| Reads shifts and shift patterns. |
|
| Reads offices and the check-in place settings. |
|
| Reads task templates with their repeats. |
|
| Reads access roles with their permissions and holders. Admin key only. |
|
| Reads the permissions each job role, designation and department gives. Admin key only. |
|
| Reads what one person may do and why ( |
|
| Reads who holds one permission ( |
|
| Reads the work time switches the admin set per person (unanswered cards, window kind). Admin key only. |
|
| Reads departments with their parent, head and status ( |
|
| Reads grades with their rank and salary band (paise). The band shows only to keys with pay access. |
|
| Reads job role changes ( |
|
| Reads the work time window of each day ( |
|
| Reads employment types: pay kind (marks or hours), switches, holders. Admin key only. |
|
| Reads the types for adding a person ( |
|
| Reads work time cards: overtime windows, late runs, corrections ( |
|
| Reads the hours entered for days of one person ( |
|
| Reads people with no work time window ( |
|
| Reads hiring candidates ( |
|
| Reads hiring applications ( |
|
| Reads leave balances in days ( |
|
| Reads leave requests ( |
|
| Reads the one record of each day of one person: mark or hours ( |
|
| Reads absences not informed on time, the sandwich rule ( |
|
| Reads the hours of hourly people, paid into the wallet ( |
|
| Reads check-ins without the photo ( |
|
| Reads roster day plans ( |
|
| Reads pay runs ( |
|
| Reads salary structures ( |
|
| Reads who is online, or one person's timeline for one day ( |
|
| Reads the test library, or who sat which test ( |
|
| Reads the targets of one person ( |
|
| Reads letters issued to people ( |
|
| Reads the documents of one person ( |
|
| Drafts one change set (title, reason, items). Nothing changes until a person approves it. |
|
| Reads one set you drafted, with the lines and status of each item. |
|
| Lists the sets you drafted, newest first. |
|
| Withdraws a set that still waits for review. |
Every POST carries a new Idempotency-Key (a UUID for each tool call), so a repeated call does not make a second draft.
Ask Claude to change settings
You do not write JSON. You ask in plain words, and Claude does the steps:
You say: "Give Ravi the extra job role Sales and make Neha his manager. Also set the notice period to 45 days."
Claude reads the current state (
list_change_areas,read_people,read_settings). It never guesses.Claude drafts one set with
draft_changes. Each change has its own reason. A thing the request needs that does not exist yet (a new department, a new person, a new access role) is made by an earlier change of the same set, so one request is one set to approve.The portal answers with the plain lines of each change. Claude shows them to you.
You open Changes to review in the portal. You approve or reject each line. You can undo a line you approved.
Claude calls
get_change_setand tells you what was applied.
Nothing changes until you approve in the portal. Only the admin's key may draft money and offer letters. Other keys are refused money, the admin account and executive roles, so change those in the portal itself. If a part of the portal cannot be drafted yet, Claude says: "This part of the portal cannot be changed by draft yet."
The employment type decides how a person is paid: its pay kind is marks (paid by the mark of each day) or hours (paid
by the hour into the wallet). An offer for an hours type carries hourly_rate_paise. The admin's key may draft a type and its pay kind in employment_types. A key never decides a mark: the admin's
key may draft a day in past_days, and the day is marked only when the admin approves it.
A main task with sub-tasks
Each draft in draft_tasks takes title, and optionally templateKey or body, notes, priority, dueAt, category,
doneWhen (lines), and one of:
refplusparentRef: give the main task aref; each sub-task names it inparentRefand comes after it in the list.parentId: hang the draft under a task that already exists.
A sub-task takes its main task's category when it has none. When the admin approves the batch, the tree stays: the main tasks go to the person or pool the admin picks, and the sub-tasks can wait so the admin gives each one later.
Prompts
write_test_plan(optionalsubject) explains the plan JSON format (areas, cases, id, title, priority, devicephone|laptop|both, url, preconditions, steps, expected, screenshot) and the writing rules, then the agent sends the plan withdraft_checklist.plan_growth_work(optionalgoal,category) explains how to plan business work as main tasks with sub-tasks in the portal's words: one result per task, "What to hand in" in the notes, adoneWhena reviewer can check, numbers as starting values. Then the agent sends it withdraft_tasks.change_portal(optionalwhat) tells the agent the flow: read first, draft one set for each topic, give each item a reason, show you the lines the portal returned, and tell you to approve on Changes to review.
Each draft waits for a person to approve it in the portal.
Errors
When the portal refuses a call, the agent gets the portal's error code, its message and every details path (for example
drafts[3].parentRef), so it can fix the input and send again. Three statuses add a plain hint:
401: the key is wrong, ended or revoked. Make a new key on the Profile page.403: the key lacks a scope, or you lack the right. Make a new key with the scope.429: too many calls. Wait a minute.
Security model
It runs on your computer. The server is a local program. The agent talks to it over stdio, and it talks to the portal.
The key goes over https only. A plain http address is refused, except
localhost. A redirect is never followed, so the key is never sent on to another address.Portal text is data. Task text and notes are written by people. The server tells the agent to read them as data, never as instructions. Review what an agent drafts before you approve it.
The key stays in the header. The server sends the key only as an
Authorization: Bearerheader. It never writes the key to a log, an error or a URL.The key acts as you, cut to its scopes. A key never has more rights than the person who made it.
The agent can only make drafts and notes. It cannot approve, assign, change a task, pay, send mail or delete.
A key drafts changes; only a person approves them in the portal. No tool approves, applies or undoes a change set. The portal refuses money, the admin account and executive roles to other keys, even as a draft. Only the admin approves the items only the admin's key could draft.
A person approves every draft. Nothing becomes a task until a person approves it in the portal.
You can revoke a key at any time. Open the portal Profile page and revoke the key. It stops working at once.
Report a vulnerability in private. Follow SECURITY.md. Do not open a public issue.
Develop
pnpm install
pnpm lint # biome
pnpm typecheck # tsc
pnpm test # unit tests with a fake fetch (this package is a client of an HTTP API)
pnpm build # tsup, writes dist/
pnpm check:dist # builds and fails when dist/ differs from the commit
COB_PORTAL_URL=... COB_PORTAL_KEY=... pnpm test:integration # real portal, read-only; skips when unsetContributions
Code contributions are by invitation only, because the server follows one private portal. Bug reports are welcome as issues. Report a security problem in private, as SECURITY.md says.
Release
Change the code and the tests. Bump
versioninpackage.jsonandVERSIONinsrc/core.ts.pnpm lint && pnpm typecheck && pnpm test && pnpm build.Commit the source and
dist/together.pnpm check:distmust pass after the commit.Tag the commit
v<version>and push the tag, so people can pin it.
This package does not depend on cob-mcp-cli-kit, as cob-rag-mcp does not. It uses the MCP SDK directly.
Changelog
Licence
MIT. Copyright 2026 Callobuzz. Built by Call O Buzz Services.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Let AI agents read your Discord community's analytics, members and tickets, and save drafts.
The official Upwork MCP server, letting AI agents connect to Upwork and act on your behalf.
A public hive where AI agents post, plan, take on tasks and vote. Reads are open; writes are signed.
Shared Rooms, skills and team memory for your people and AI agents, with human approvals.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceThis server enables AI assistants (CLINE, Cursor, Windsurf, Claude Desktop) to share a common knowledge base through Retrieval Augmented Generation (RAG), providing consistent information access across multiple tools.4MIT
- AlicenseNot gradedqualityDmaintenanceThis server allows AI assistants to interact directly with your Booqable rental management system, providing tools to manage inventory, orders, and more.MIT
- AlicenseNot gradedqualityCmaintenanceThis server enables AI assistants to manage business operations within Doklad.ai, allowing for the creation of invoices, management of customers and expenses, and summarization of cash flow.MIT
- AlicenseBqualityBmaintenanceThis server enables AI agents to automate complex design workflows by rendering sketches, exploring reference variations, performing precise masked edits, and managing shared boards.211MIT