Skip to content
Tallyfy

MCP server

Tallyfy MCP server

Tallyfy’s MCP (Model Context Protocol) Server lets you control workflows with plain English through any AI platform that supports MCP. You describe what you want, and the AI handles the rest, no API knowledge required. For a simpler approach that uses your existing AI subscriptions without technical setup, see our BYO AI (Bring Your Own AI) integration.

What is MCP?

MCP is a universal translator between AI assistants and your business tools. It’s an open standard - governed by the Agentic AI Foundation under the Linux Foundation - that creates secure connections between AI models and external systems while you stay in control of permissions. All major AI providers (OpenAI, Anthropic, Google, Microsoft) support it.

How it works

The MCP Server exposes 107 tools that connect AI assistants (Claude, ChatGPT, Gemini, Copilot, and others) directly to your Tallyfy organization. Connect with OAuth 2.1 authentication or an API token, and you can:

  • Ask questions about your processes and tasks in plain English
  • Create, complete, and manage tasks through conversation
  • Launch and manage processes directly
  • Analyze templates and get improvement suggestions
  • Search across your organization’s workflows
  • Manage users and guests

Example interactions:

  • “Show me all overdue tasks for John”
  • “Launch the Employee Onboarding process for Jane Doe”
  • “What templates do we have for customer onboarding?”
  • “Assess the health of our ‘Employee Onboarding’ template”

What the MCP server can do

The server exposes tools across these capability areas, all callable from any MCP-compatible AI:

Search and discovery

Find tasks, processes, templates, snippets, or do a unified search across all entity types in plain English.

  • “Find all tasks related to budget approval”
  • “Search for processes containing ‘customer onboarding’”
  • “Show me templates for HR workflows”

User and access management

List and manage members, guests, groups, and roles. Invite new users, change roles, enable or disable accounts.

Task management

View your own tasks, list tasks assigned to teammates, see all tasks within a process, create standalone tasks from natural language, complete or reopen tasks, update titles and deadlines, manage comments, and report or resolve task issues.

Process management

List running processes, launch new ones from a template, view or update process state, archive completed processes, and reactivate archived ones.

Template design and optimization

Get any template’s detail or a full inventory, view steps, run a health assessment, see step dependencies and visibility conditions, get deadline recommendations, modify template properties, clone templates, and add or reorder steps.

Form field management

Add form fields to steps, update properties, move fields between steps, reorder fields, delete fields, get AI suggestions for useful fields, and manage dropdown options.

Kickoff fields

View, add, update, reorder, and delete kickoff (prerun) fields on templates. Get dropdown/radio/multiselect options for kickoff fields. Get AI recommendations for process launch form fields, complete or reopen the kickoff form on a process.

Automation and workflow logic

Create, update, and delete automation rules. Review all rules with duplicate detection, get consolidation recommendations that flag orphaned rules and merge candidates, and inspect step visibility conditions.

Organization tools

Manage tags, folders, and tags on templates and processes.

AI template generation

Generate template drafts from a prompt, an uploaded document (PDF, DOCX, or image), or convert a document into structured Tallyfy steps and form fields. Suggests template names and procedure steps for fast bootstrapping.

Security and authentication

The MCP Server uses OAuth 2.1 with PKCE for secure authentication. The well-known endpoint at /.well-known/oauth-authorization-server (per RFC 8414) lets AI clients discover authorization endpoints automatically. OAuth scopes use dot notation - for example, mcp.tasks.read, mcp.processes.write, mcp.templates.read.

Key security features:

  • Only authorized users can access your organization’s data
  • Every action respects your existing permissions
  • Audit trails track who did what
  • Dynamic Client Registration (RFC 7591) with redirect URI validation
  • Refresh token rotation with reuse detection
  • Authorize button stays disabled until you check at least one scope
  • Successful deep-link consent auto-closes the consent window so you return to your AI client without manual action
  • Cancelling consent returns a clean error page instead of raw protocol output

Available MCP integrations

Pick your AI platform:

  • Claude (Anthropic) - Remote MCP for Claude Desktop, claude.ai connectors, and Claude Code on Pro, Max, Team, and Enterprise plans
  • ChatGPT (OpenAI) - Custom Connector + ChatGPT Apps on Pro/Plus/Team/Enterprise/Education plans
  • Google Gemini - Gemini CLI MCP config and Gemini Enterprise custom MCP data store
  • Microsoft Copilot Studio - Custom MCP connector, Power Platform certified path
  • Cursor, Cline, Continue - Standard MCP server config in your .mcp.json or workspace settings
  • MCP Inspector and CLI tools - Connect via the standard streamable-http transport

Future capabilities

Coming soon:

  • Advanced reports with charts and insights
  • Connections to other middleware platforms
  • Real-time team collaboration features

Recent updates

  • OWASP MCP Top-10 audit closed (May 2026) - the server passed a full OWASP MCP-specific threat model with hardening for prompt injection, indirect tool calls, scope confusion, and token leakage paths
  • Cloud Run mirror operational (May 2026) - https://mcp-gcp.tallyfy.com/ brought up as the Tier-1 origin for Gemini Enterprise customers; the Google Cloud build pipeline is now unblocked and deploys on every push
  • Cross-process structured logging - the MCP host’s request/response audit trail now flows into Tallyfy’s central logging backend, joining the same observability stack as API and email events

Technical architecture

The MCP server is built from:

  • Core SDK - Python client that calls Tallyfy’s API
  • MCP server - The protocol implementation, built on FastMCP, transport is streamable-http
  • Tool system - Structured tool definitions that tell AI clients what actions are available, with ToolAnnotations (read-only or destructive hints) on every tool
  • OAuth 2.1 layer - Dynamic Client Registration (RFC 7591), PKCE S256, RS256-signed JWT access tokens with refresh-token rotation

Public endpoints:

  • https://mcp.tallyfy.com/ - the primary streamable-http MCP endpoint (DigitalOcean)
  • https://mcp-gcp.tallyfy.com/ - the Google Cloud Run mirror (Tier-1 path for Gemini Enterprise)
  • https://mcp.tallyfy.com/.well-known/oauth-protected-resource - OAuth resource metadata
  • https://mcp.tallyfy.com/.well-known/oauth-authorization-server - OAuth authorization server metadata
  • https://mcp.tallyfy.com/.well-known/jwks.json - public RS256 signing key
  • https://mcp.tallyfy.com/.well-known/openai-apps-challenge - OpenAI Apps directory verification challenge
  • https://mcp.tallyfy.com/health - health check
  • https://mcp.tallyfy.com/privacy - MCP-specific privacy summary

Both endpoints share the same OAuth backend, the same 107 tools, and the same Tallyfy organization data. Pick mcp.tallyfy.com by default; use mcp-gcp.tallyfy.com when your AI platform (Gemini Enterprise in particular) prefers a Google Cloud-hosted origin.

Limitations of text-based AI interfaces

MCP is powerful, but text-based AI chat has real limitations when you’re working with Tallyfy’s visual workflow features:

Visual interface constraints

  • No process tracker: Can’t display Tallyfy’s bird’s-eye view of running processes with progress indicators
  • No template builder: Can’t show the drag-and-drop interface or visualize step dependencies
  • No aggregated views: Can’t present filterable views of processes and tasks at a glance
  • No live updates: Can’t show real-time progress as team members complete tasks

Form field challenges

  • Dropdowns: Text-based AI must list all options as plain text, making selection tedious
  • Multi-select inputs: Form inputs lose their visual interface when reduced to text
  • Date/time pickers: Calendar and time tools become manual text entry
  • File uploads: Can’t handle file uploads through conversation

Assignment and collaboration barriers

  • Assignee selection: Can’t show Tallyfy’s visual suggestions for members, guests, job titles, and groups
  • Bulk operations: Managing multiple assignments through individual text commands is slow
  • Collaboration: Real-time features like simultaneous template editing become sequential

Template creation limitations

  • Step visualization: Can’t display all steps at once or show their relationships visually
  • Reordering: No drag-and-drop - must use text commands
  • Automation setup: Complex automation rules are harder to configure through conversation
  • Preview: Can’t visually preview how a template will work before deploying it

Where MCP works best

Text-based MCP shines in specific scenarios:

Search and discovery

Ask specific questions, get targeted answers:

  • “Find all templates related to employee onboarding that include background check steps”
  • “Show me all customer onboarding processes that took longer than 5 days”
  • “Which tasks are overdue and assigned to the sales team?”
  • “Find templates that use the ‘Budget Approval’ form field”

Template generation from documents

Got a messy document? AI turns it into structured templates:

  • Form field creation: Upload forms and let AI generate form fields with validation rules
  • Flowchart conversion: Turn process diagrams into templates with proper step sequences
  • Automation generation: Convert plain-language business rules into automation rules
  • Bulk field creation: Generate multiple related form fields from document analysis

What-if scenario testing

  • “If I set up an automation to route tasks by deal value, how would it handle these 5 example deals?”
  • “What would happen if we removed the approval step for purchases under $1,000?”
  • “How would task distribution change if we reassigned John’s templates to Sarah?”
  • “Based on historical data, estimate completion time if we add a review step”

Process updates from changed documents

Compliance docs changed? AI spots differences and updates everything:

  • “Here’s our updated SOX compliance procedure. Update our audit template to match”
  • AI identifies exactly what changed and updates only affected parts
  • Compare old and new documents to generate precise template modifications
  • Apply consistent changes across multiple templates from policy updates

Pattern recognition and optimization

  • “Look at one-off tasks added to hiring processes last month - which should become permanent template steps?”
  • “Which steps consistently cause delays?”
  • “Find similar templates that could be merged into one”
  • “Identify patterns in high-performing processes and suggest improvements”

Reporting with citations

  • “Which step in our sales process has the longest average completion time?”
  • “Show all instances where required approvals were skipped, with links to the processes”
  • “How has customer onboarding time changed over the last 6 months?”
  • “Which departments most frequently add ad-hoc tasks?”

What it looks like in Claude

Real screenshots from claude.ai with the Tallyfy connector active. The full set with longer captions lives in the Claude integration guide - here are four highlights.

Building a workflow from a diagram. Drop in a BPMN flowchart, get a working Tallyfy template with steps and conditional automation rules.

Claude creating a 7-step Business Trip Request template in Tallyfy with 5 automation rules matching a BPMN diagram

Confirming scope before pulling data. Claude uses the ask_user_question MCP tool to clarify ambiguous requests before any tools fire.

Claude.ai Cowork sidebar showing the Tallyfy connector connected, with Claude asking four clarifying questions before querying pending tasks

Reviewing an existing process. Asked to review a Client Onboarding workflow, Claude confirms whether you want suggestions only, edit-on-approval, or direct edits.

Claude.ai Cowork sidebar reviewing a Client Onboarding process and confirming desired scope via a four-option picker

Multi-tool analysis with progress tracking. Claude chains get_me, get_my_tasks, and counting/verification logic. The result: 822 open tasks, 805 explicitly overdue, zero due in the next 30 days because everything is already past its date.

Claude.ai answering a pending-tasks question against the Tallyfy MCP server, with the progress sidebar showing four completed sub-steps and a detailed breakdown of 822 open tasks

Why Tallyfy is the AI control layer

What Tallyfy doesWhy it matters in the age of AI
Define process stepsYou can’t automate anything without a recipe.
Hand small tasks to AIMassive reduction in mistakes, omissions, and hallucination.
People approve workAccountability. You can’t blame AI for mistakes.
Hybrid people + AI tasksAI can’t do every task in a process.
Track real-time statusTracking AI sessions at scale is a nightmare.
Gradually shift tasks to AIA total re-do will break a process that works today.
Get in touch
About Tallyfy