Founders OS is an open-source MCP (Model Context Protocol) server for startup and small-business founders. Its goal is to expose the complete business context founders need in their daily work—CRM, projects, tasks, finance, feeds, semantic memory, and playbooks—to Claude, Cursor, or any MCP-compatible AI client, making all business data accessible through a single connection. The project is developed by OurThinkTank and is licensed under MIT.
Purpose and overall structure Founders OS contains 12 modules and 94 tools, covering customer relationships, tasks, projects, playbooks, tags, finance, feeds, memory, cross-domain views, member directory, audit and recovery, and diagnostics and meta-information. It requires users to provide their own Supabase project, an embedding API key (OpenAI by default), and an MCP-compatible AI client.
Main modules and capabilities CRM and pipeline management: Customers are organizations, and contacts are individuals; the two are always recorded separately, so a contact can be moved under a different customer without losing history. Pipeline stages are prospect → lead → opportunity → customer → renewal, plus churned and inactive.
Tasks: Unified task management, supporting both organization and individual scopes, entity associations, AI assignment, task dependencies, progress notes, and a task-to-memory bridge. You can use @claude or @gpt as the assignee, and get_task_summary presents a dedicated AI work queue. Setting blocked_by_task_id establishes a dependency, and after completing a blocking task, the response will indicate which tasks have been unblocked. When completing a task, setting store_as_memory=true persists the completion note as organization-level memory. Tasks can be linked via task_links to any entity type, such as customers, contacts, interactions, deals, projects, playbooks, and memories.
Projects: A project is an entity whose primary record is a project tag (such as #acme-rebuild). get_project returns the project card, status, associated tags, recent tasks grouped by status, and customers marked as belonging to the project.
Playbooks: Reusable named orchestration templates. A playbook is defined once, run against a customer or another subject, andYou can create Founders OS native tasks; when the corresponding MCP tools are connected, external actions are also triggered, such as creating a GitHub repository, posting a Slack message, or scheduling a calendar event. If a connector is unavailable, the step gracefully degrades to a task with a [manual] tag, so the playbook can still run. Step types are divided into native_task and external_action. Supported placeholders include {{customer.name}}, {{customer.slug}}, {{playbook.start_date}}, {{playbook.start_date+Nd}}, {{contact.primary.name}}, {{memory:key}}, resolved at runtime.
Tags: a shared tag registry with soft validation. Tags are advisory: unrecognized tags emit a warning but never block an action, and new tags are automatically registered on first use. The conventions are #project-name for projects, @person for people, and !state for meta-states (such as !needs-review).
Finance: simple double-entry bookkeeping partitioned by FOUNDERS_OS_COMPANY_ID, with per-user access control, allowing the company ledger to be opened only to specific teammates. Support for setting different FOUNDERS_OS_COMPANY_ID values to isolate multiple ledgers. By default the company owner has financial access; others must be explicitly granted via set_financial_access.
Feeds: a built-in subscription reader supporting RSS, Atom, and JSON Feed, stored using Postgres, with subscriptions, digests, search, favorites, and pinning. Categories include tech, startups, business, finance, product, design, engineering, ai, crypto, science, news, personal, other.
Memory: pgvector-based semantic memory, supporting both personal and organizational scopes, near-duplicate detection, metadata filtering, and pagination. memory_recall supports min_score, source_tool, created_after, created_before, offset, limit, project, scope and other filter conditions. On write, it checks whether cosine similarity reaches or exceeds 0.92 and gives a conflict prompt, with the option to force storage or skip. The embedding provider is set via EMBEDDING_PROVIDER; before running 002_memory_schema.sql, EMBEDDING_DIM must be set to match the model. Dimensions are permanent, and changing providers later requires re-embedding the memory table.
Other modules: Surfaces provides cross-domain read-only views, combining data from tasks, CRM, finance, and feeds into dashboards that can be rendered directly; Members is the organizational member directory, mapping FOUNDERS_OS_USER_ID slugs to display names and marking the company owner; Audit and Restore provides structured audit logs and soft-delete recovery; Diagnostic and Meta provides diagnostic and meta-information tools.
Rendering Contract Tools with rendering include a render field, using a four-level ladder that lets AI clients choose the most visual output they support: visual primitive tools (artifact/widget/canvas), inline rich output (HTML/SVG/JSX), markdown tables, and finally plain text. The contract is delivered through three channels in order of attention strength: the server instructions field at MCP registration, get_session_start.rendering_contract, and the rendering_contract reminder in every response. The canonical source is located at packages/mcp-server/src/contract.ts, and any MCP client can get rich rendering without a plugin. The current contract_version is 4, and on mismatch a contract_version_warning appears on get_session_start and get_version.
Installation and Usage You need to first create a Supabase project and run supabase/setup.sql in the SQL editor. This file builds the complete schema from scratch, including extensions (vector, uuid-ossp, pg_trgm), all tables, indexes, RLS policies, functions, views, maintenance tasks, and the Supabase pair for 2026-05-30 andProjects created afterward are removed from automatic default permissions, a change needed to be compatible with the Data API authorization. If you do not use the default dimension, you can use the wizard at foundersmcp.com/setup, which generates the same SQL that matches the embedding dimension to the selected provider.
The MCP server runs via npx. All clients (Claude Desktop, Cowork, Cursor, Continue.dev, Zed, or any spec-compliant MCP client) use the same configuration. The fastest way is to use the foundersmcp.com/setup wizard: after entering Supabase and embedding credentials, it generates a filled-in configuration for copying or downloading, and the credentials never leave the browser. You can also manually paste the configuration block into the client's mcp.json (in Claude Desktop, that is the configured MCP servers section).
First-run guidance When the database is empty, the tools include guidance prompts in their responses. get_session_start detects a fresh installation and suggests a guided walkthrough: add the first customer, create the first task, and optionally set up a financial account. You can also have the AI run "run the welcome demo" or "show me the conflict-resolution walkthrough"; list_demos returns the built-in interactive scripts.
Limitations and current status Aside from the installation requirements above, information about fees, pricing, platform support scope, specific effects, and similar details is not currently provided. The project is under the MIT license; external contributions are not currently accepted, but issue reports are welcome on GitHub, and security issues should be reported through the SECURITY.md channel.