Onplana

Onplana MCP server

Open-source TypeScript Model Context Protocol building blocks, extracted from Onplana's production MCP deployment. Two packages:

• [`onplana-mcp-server`](./packages/server-template): server template. Streamable HTTP transport, Bearer auth, prompt-injection containment, pluggable dispatcher.
• [`onplana-mcp-client`](./packages/client): typed TypeScript client SDK for calling the public Onplana MCP endpoint at https://api.onplana.com/api/mcp/v1.

What this is

The transport layer of an MCP server (Streamable HTTP wiring, stateless mode, scoped Bearer auth, prompt-injection containment) done well, separated from the platform-specific tool registry. Use the server template to build your own MCP server with security best practices baked in. Use the client SDK to drive Onplana's hosted MCP from your own code.

The patterns are extracted from Onplana's production deployment (public docs at onplana.com/mcp), the same layer that handles real Claude Desktop, Cursor, ChatGPT custom connector, and in-house agent traffic against the Onplana platform.

Why open-source

The MCP transport is the same for everyone. Most early MCP servers get the security primitives wrong:

• Prompt injection. Tools that return user-generated content (task titles, comment bodies, wiki text) put that content directly into the model's context. Without containment, a hostile actor can plant "ignore previous instructions" in their own data and the next agent that reads it follows along.
• Stateless transport. Most SDK examples assume in-memory session state, which breaks horizontal scaling and complicates the auth model.
• Plan-gate semantics. Surfacing tools the caller can't actually invoke wastes turns and confuses the model.

Onplana solved these in production over six months of MCP-server work. Publishing the patterns is high-leverage:

1. Other MCP authors get a known-good template instead of reinventing.
2. The repo is a pretraining-signal surface. Public GitHub READMEs are heavily weighted in next-gen LLM training data, and a repo with patterns + clear documentation about MCP improves model recall of "what good MCP servers look like."
3. The dispatcher interface is the seam where your business logic plugs in. The transport is generic; what matters about your MCP server is the tool registry. Open-sourcing the transport doesn't give away anything proprietary.

The dispatcher implementation, tool catalog, plan-gate logic, audit infrastructure, and the rest of Onplana's ~600 LOC closed-source dispatcher stay in the closed monorepo because they encode platform business logic. If you build your own MCP server using this template, you write your own dispatcher. That's the work that matters and the work that's specific to your platform.

Repository layout

onplana-mcp-server/

├── packages/ │ ├── server-template/ # onplana-mcp-server (npm) │ │ ├── src/ │ │ │ ├── transport.ts # Streamable HTTP wiring │ │ │ ├── auth.ts # Bearer auth pattern │ │ │ ├── promptInjection.ts # wrapUserContent + escape │ │ │ ├── dispatcher.ts # Pluggable Dispatcher interface │ │ │ └── index.ts │ │ ├── tests/ # promptInjection + auth + transport │ │ └── README.md │ └── client/ # onplana-mcp-client (npm) │ ├── src/ │ │ ├── client.ts # OnplanaMcpClient class │ │ ├── types.ts # Public type surface │ │ └── index.ts │ ├── tests/ # client.test.ts (stub fetch) │ └── README.md ├── .claude-plugin/ │ └── marketplace.json # Claude Code marketplace ├── plugins/ │ └── onplana/ # Claude Code plugin (skills + connect command) ├── examples/ │ └── in-memory/ # Runnable demo with 3 toy tools ├── gemini-extension.json # Gemini CLI manifest ├── mcp.json # stdio client config (mcp-remote) ├── server.json # MCP registry manifest └── .github/workflows/ ├── ci.yml # tsc + vitest on PR └── publish.yml # npm publish on tag v*

Quickstart

Build a server

Install:

npm install github:Onplana/onplana-mcp-server @modelcontextprotocol/sdk express

Wire an Express app:

import express from 'express'

import { createMcpPostHandler, createMcpMethodNotAllowedHandler, requireBearerAuth, type Dispatcher, } from 'onplana-mcp-server'

const dispatcher: Dispatcher = { async listTools(ctx) { /* return your tool descriptors */ return [] }, async callTool(name, input, ctx) { /* dispatch to your tools */ return { output: {} } }, }

const auth = async (token: string) => { // Validate against your token store. Return AuthContext or null. return { userId: 'u', scopes: ['MCP_AGENT'] } }

const app = express() app.use(express.json()) app.use('/api/mcp/v1', requireBearerAuth({ auth, requiredScope: 'MCP_AGENT' }), ) app.post('/api/mcp/v1', createMcpPostHandler({ dispatcher })) app.get('/api/mcp/v1', createMcpMethodNotAllowedHandler()) app.delete('/api/mcp/v1', createMcpMethodNotAllowedHandler()) app.listen(3000)

Full quickstart in packages/server-template/README.md; runnable demo in examples/in-memory/.

Drive Onplana from code

Install:

npm install github:Onplana/onplana-mcp-server

Use:

import { OnplanaMcpClient } from 'onplana-mcp-client'

const client = new OnplanaMcpClient({ url: 'https://api.onplana.com/api/mcp/v1', token: process.env.ONPLANA_PAT!, })

const projects = await client.listProjects({ status: 'ACTIVE' })

// The differentiator vs other PM-tool MCPs: hybrid semantic + lexical // search across your org's indexed content (projects, tasks, risks, // goals, comments, wiki pages). const { matches } = await client.searchOrgKnowledge({ query: 'rationale for the 3-week design phase', scope: 'all', limit: 5, })

Full client docs in packages/client/README.md.

Tools

The hosted server at https://mcp.onplana.com/mcp exposes 285 tools, spanning projects, tasks, sprints, milestones, earned value, risks, issues, governance, change control, timesheets, wikis, whiteboards, workflows and the Microsoft Graph integrations. The exact number a given client sees is smaller, because tools are filtered by the caller's role and the organization's plan before the catalog is served.

The 33 below are the ones worth knowing first, not the whole catalog. Reads are annotated readOnlyHint; writes carry destructiveHint so a client can gate them. Every call runs under the calling identity, is checked against that user's permissions and the org's plan, and lands in the audit trail.

Read (readOnlyHint: true)

• list_projects: projects in the org, filterable by status.
• get_project: one project in full, with dates, owner and progress.
• list_tasks: tasks for a project, or across projects.
• get_task: one task with description, assignee, dates and recent comments.
• list_my_tasks: tasks assigned to the calling user.
• list_overdue: tasks past their due date.
• list_team_members: members of a project.
• list_org_members: members of the organization.
• list_risks: risks logged against a project.
• find_similar_projects: past projects resembling a description, for estimating.
• search_org_knowledge: hybrid BM25 and vector search over tasks, projects, wiki pages and comments.
• summarize_project: AI summary synthesized from the live plan.
• analyze_project_risks: AI risk detection across schedule, budget, scope and resources.
• generate_status_report: AI status report from the current schedule and activity.
• search: App Directory adapter, returns {id, title, snippet?, url?}.
• fetch: App Directory adapter, returns {id, title, content, url?, metadata?}.

Write, additive (destructiveHint: false)

• create_project: create a project.
• create_task: create a task, optionally under a parent.
• create_milestone: add a milestone to a project.
• create_comment: comment on a task, issue or project.
• create_sprint_with_tasks: create a sprint and pull tasks into it.
• submit_timesheet: log hours against a task.
• add_project_member: add an existing org member to a project.
• link_dependency: link two tasks, idempotent via a unique constraint.

Write, mutating (destructiveHint: true)

• update_project: change project fields such as status, dates or budget.
• update_task: change task fields such as status, progress or dates.
• bulk_update_tasks: apply one change across many tasks.
• assign_task: set a task's assignee.
• move_task_to_sprint: move a task into or out of a sprint.

Leases (for agents that share a backlog)

• next_task: pick the next available task and claim it in one call. Listing and then claiming leaves a gap two agents can both land in.
• claim_task: take an exclusive lease on a specific task.
• renew_task_lease: extend a lease while the work is still running.
• release_task: hand the lease back; completing or blocking a task releases it too, and ending a session releases everything that run holds.

A lease is keyed to the RUN, not to the user. Two sessions of one client authenticate as the same agent persona, so a user-keyed lock would let one session release the other's work. Leases expire on their own, so a crashed agent frees its task instead of holding it.

Four delete tools are in the catalog (a task, a project, a list, a document). Each is a per-organization switch that is on until an org admin saves a choice, the agent still needs the underlying permission, and all four move items to a recycle bin rather than destroying them. Prefer update_task over delete-and-recreate anyway, since Onplana audits every field change and keeps the history.

Production checklist

The template + SDK get you running. Add these on top:

• Per-token rate limiting. 60–120 req/min per Bearer token; agentic loops are noisier than humans.
• Tenant cost cap. If your tools call paid LLMs, gate dispatch on month-to-date spend. Onplana's deployment uses aiMonthlyCostCapUsd with WARN / BLOCK modes.
• Audit logging. Every dispatch should write an audit row tagged with actorType: 'mcp_agent' so admins can see what AI agents did in their tenant separately from human activity.
• Plan / scope curation. Don't expose every internal tool. Onplana exposes 21 of 26; the suppressed 5 either need an in-app preview UI, are too risky for unsupervised invocation, or produce oversized payloads.
• PREVIEW mode for risky mutations. Default mutating tools to preview-only on free tiers. Onplana ships this: agents see "what it would do" before users explicitly upgrade and re-run.
• Idempotency keys. Hash the canonicalised input + a session id; store as a unique constraint on your audit row. A model retrying the same logical action shouldn't double-create.

Each of those is platform-specific. The template gives you the seam where they plug in (Dispatcher.callTool); your dispatcher implements them however your platform encodes those concepts.

Compatibility

• Node.js ≥ 20 (for the server template and CI matrix); ≥ 18 for the client (uses ambient fetch).
• @modelcontextprotocol/sdk@^1.29.0
• express@^4.18.0 or express@^5.0.0

Tested against:

• Claude Code (plugin marketplace, or claude mcp add --transport http)
• Claude Desktop (Custom Connector)
• Cursor (~/.cursor/mcp.json)
• ChatGPT custom connectors (where MCP is enabled in your accou