feedback: Pure MCP for 0$ cost

[uros-mil]

Feedback Summary

Extend task-master MCP server to provide prompts for all the actions (like parse-prd) and use other/existing MCP for local file access/update, access to PRD source, and web-search.

User Context

I'm using Claude Desktop for more and more engineer tasks, including PRD/codebase analysis, class/flow diagram generation from codebase. Currently my PRD analysis is combination of manual steps planning and AI-assisted codebase and documentation analysis.

I would like to use task-master to improve my workflow by automating PRD parsing and docs/codebase matching, without unnecessary cost increase. Cloude Desktop could already offer LLM calls. Similarly Perplexity/Brave/Tavily or Sequential-Thinking are available through MCP servers.

User Impact

• Cost reduction in LLM usage as monthly subscription already covers it
• No extra work to integrate ADO/Jira/Linear/GitHub/Gitlab, Notion/GDocs, Brave/Tavily, and similar, as those (and many others) are automatically available through existing MCP servers.

[fourcolors]

in addition to this, claude code as well.

[eyaltoledano]

This requires either sampling (which to my knowledge is not available in most AI IDEs) or an MCP to MCP protocol, both of which are not currently within reach.

Open to research and PRs

[nnWhisperer]

Hello, how is cost reduction taking place by using Claude Code and MCP @uros-mil ? In addition to the coding assistant, does Claude Desktop(subscription) give you api keys that you can use on projects?

[uros-mil]

Hi folks,

I was thinking about this and have a couple of ideas.

Sampling support in MCP Client

I do agree that this will be the best approach if possible, as it will require minimal changes to the current workflow.

But this is not going to be coming soon, since most of the platforms charge per month and this will impact negatively their formula for average requests per customer.

Prompts from MCP Server

My idea was to use Prompt support in MCP protocol to provide MCP Clients with a user+system prompt (as prompt-template) you are already using for actions:

• parse-prd: https://github.com/eyaltoledano/claude-task-master/blob/09d839fff5c3800075b921d8569bee7fe6143eb5/scripts/modules/task-manager/parse-prd.js#L152
• complexity: https://github.com/eyaltoledano/claude-task-master/blob/09d839fff5c3800075b921d8569bee7fe6143eb5/scripts/modules/task-manager/analyze-task-complexity.js#L19

Those will need some adaptation, so users could easily reference a prompt-template exposed in the MCP Client in their message/reply. But code reuse between prompt-templates exposed to the MCP Client and internal usage should be possible.

Prompt generation similar to Repo-prompt (https://repoprompt.com)

Another idea is to adapt your MPC Server to return full prompt content to the MCP Client, which users could then copy/paste inside the input field.

Repo-prompt is doing the following things to reduce cost for the users:

• They create a full prompt with all necessary content
• then expose it with the Copy button
• User then pastes the whole thing to Claude Desktop (or any other AI assistant) and executes a request to model and receive the complete result

Example of MCP data-flow

1. The user writes a message that references their PRD file and asks for a task creation plan to be created
   • A path could be provided to taskmaster-ai, which reads the file content
   • or the file could be read by some other tool (like desktop-commander) and the content provided to the taskmaster-ai tool
2. task-master creates instructions from PRD content, and the existing user+system prompt
3. This instruction is returned to the MCP Client, which
   • ideally execute it - it's an open question how to format the response so MCP Client does execute it, since they are not going to blindly execute prompt-tempalate returned from any MCP Server (that would make it sampling)
   • or just shows generated instruction as an artifact, which the user could copy/paste as a new message
4. The received task plan could then be stored on the file system with other MCP servers (like desktop-commander)
5. The process will continue, in the same way, to generate a single task file and start implementation on it

The same flow could be adapted to the majority of taskmaster-ai tools.

[Crunchyman-ralph]

if cost is your worry, it shouldn't be, its more expensive to sample cursor, and we will eat up all your credits with sampling.

If the reason you want to sample is for something other than cost, what is it ? Ask yourself what problem are you trying to solve other than the price. If its just the price, its not worth it

[dantra13]

@Crunchyman-ralph Since you mentioned the price, do you have an estimate of how much it costs to use all the APIs? It depends on how it's used, but it's almost my first time using external services via API.

[Crunchyman-ralph]

5$ or 10$ if you use it a lot

[uros-mil]

Regarding price:

• It is one of the points
• When you have a paid subscription to Claude Max, you prefer not to spend extra
• People also get access to the Claude subscription from their employer, but do not have access to API keys (or have limited credit)
• Not thinking about the price motivates people to use the tool much more often

Regarding workflow:

• I would prefer to have TaskMaster tools as Claude's custom commands
• I'm experimenting with this, and if I find a simple way to refactor code in claude-task-master/scripts/modules/task-manager, I'll share it
  • The idea is to extract the currently used prompts into template files
  • This way current code will work the same way - it will read prompts from a template file instead of them being hard-coded internally
  • We could have an automated way to create MD files for Claude Code from those same templates
  • This could simplify further experimentation when combining with other MCP tools - e.g., to load various docs/specs from Web/Notion or inject files/codemap from repo

By the way, I want to congratulate the TaskMaster team on doing great work. 👏

Sorry for the late answers. I'm quite busy at the moment. Expect answers in 5-7 days.

[Crunchyman-ralph]

@uros-mil these are all great points, and as task-master grows, we realize yes the prompts are important but it is becoming much bigger than that.

I have yet to see a lot of adoption from claude code, we usually see cursor, windsurf, cline, roo, not much claude code.

But I believe claude code has some really good mcp client integrations, and in that specific case, we could use something called sampling!

Definitely something to consider, and we are going to keep this open as we wait for more people to give their feedback on this!

[neno-is-ooo]

see #601 :)

[Crunchyman-ralph]

We have implemented sampling in 0.20.0, you can find it by running task-master models --setup and searching for "mcp"