Canh AI

The description states that canh is a hosted, multi-user task and project layer exposed through the Model Context Protocol (MCP). It supports:

• Creation, finding, updating, assigning, organizing, and discussing tasks.
• Idempotent ingestion and persistent history to avoid duplication.
• Collaboration across teams and organizations.
• A web control surface for sign-in, teams, credentials, and account management.
• Support for boards, due dates, recurrence, templates, comments, task relationships, assignment, cross-team moves, event history, recoverable deletion, and protected batch operations.

It is described as model-agnostic, meaning it does not run an LLM itself and stores no source system credentials. The user’s chosen assistant calls canh over MCP to perform actions.

Inference This appears to be a backend infrastructure layer designed to support AI agents working with structured task data, rather than a direct-to-user interface.

The author claims that canh is:

• A “to-do list you never open”, implying it works in the background.
• A durable, shared MCP task layer for both people and AI agents.
• Designed to preserve context across conversations, avoiding re-creation of tasks.
• Built to separate task management from company systems of record (e.g., Jira).
• Intended as a “sustainable SaaS product” with potential for billing and operational tooling.

The positioning evolved during OpenAI Build Week, where the author shifted focus toward:

• Agent-agnostic repository guidance.
• Richer event history, comments, and recoverable deletion.
• Expanded team collaboration and lifecycle controls.
• Improved concurrency and authorization models.

Inference The product evolved from a personal prototype into what the author describes as a foundational SaaS layer for LLM-based task management.

The description does not clearly define a specific customer segment or ICP. It mentions:

• People using canh individually or with teams and organizations.
• AI agents that interact with the system via MCP.
• Users who want to extract action items from email, Slack, and meeting notes.

It also implies that users may be those who do not want to use traditional task managers like Jira or Trello for small follow-ups or personal tasks.

Inference The target is likely early adopters of AI tools who are looking for a lightweight, structured way to manage work that integrates with their existing LLM workflows.

There is no evidence in the description of pricing models, monetization strategies, or any commercial structure. The author mentions:

• “Next I want to add production billing and operational tooling.”
• “canh is a live, end-to-end SaaS foundation.”

But no details about how users will pay, what pricing tiers exist, or if there are paid features.

Inference The business model is not yet defined, though the author intends to build one around SaaS.

The system uses:

• Backend: Python 3.12, Starlette, FastMCP, async SQLAlchemy, PostgreSQL 18, Alembic.
• Deployment: Google Cloud Run and Cloud SQL.
• Authentication & Web UI: Firebase Authentication and Hosting.
• Infrastructure Management: Terraform.

It is built with:

• Strict tenant isolation.
• Scoped credentials.
• Idempotent writes.
• Explicit actor attribution.
• Optimistic concurrency.
• Immutable batch identity.
• Compare-and-swap protection for batch undo.

The author also mentions using AI agents to test scenarios and implement features, iterating through at least four cycles of development.

Inference The technical stack is mature enough to support multi-tenant, concurrent operations with strong consistency guarantees. The use of AI in development suggests a high degree of automation and iterative refinement.

The description states:

• Canh is a live, end-to-end SaaS foundation.
• It has a production MCP endpoint.
• A web management surface exists.
• There is a comprehensive automated test suite (648 Python tests, 22 JavaScript tests).
• The system supports multi-tenant collaboration.

However, there is no evidence of actual users, revenue, or adoption beyond the author’s own development efforts.

Inference The product shows technical maturity and readiness for deployment, but lacks commercial traction or user feedback.

The author explicitly contrasts canh with:

• Jira: For enterprise-level systems of record.
• Trello: For traditional task managers.
• Other PM tools that are designed around people opening apps rather than LLMs reliably reading and updating tasks.

It is positioned as a layer between AI agents and structured work, separate from existing tools.

Inference Canh targets a niche within the AI agent workflow space, where existing tools are not optimized for LLM interaction.

• No commercial traction or revenue: The product is described as a prototype that has been extended but lacks any evidence of users or monetization.
• Unproven market demand: There is no indication that the target audience (AI agents and users) actually needs this specific solution.
• High technical complexity without user validation: Features like batch undo, cross-team moves, and concurrency control are advanced, but without real-world usage, their value is unclear.
• Founder-only team: Only one person is listed as part of the team, which may limit execution speed or scalability.

Inference The risk of misalignment between product development and actual market needs is high.

1. What specific use cases have you identified for AI agents interacting with this system?
2. Have you conducted any user research or interviews with potential customers?
3. How do you plan to monetize the platform? Are there any early adopters or pilot programs?
4. What are the key assumptions about how LLMs will interact with task systems, and how have these been validated?
5. What is your roadmap for integrating with existing tools (e.g., Slack, email clients)?
6. How do you intend to scale beyond a single developer’s capacity?