Description
Describe the bug
Agent Mode in VS Code fails to create a pull request because it skips the get_me prerequisite call. It guessed the repository owner from the local path (dimitriosphilliou
- my full name) instead of using the authenticated GitHub handle (D1M1TR10S), so Agent Mode returns 404 when trying open the PR.
This may need some work on Agent Mode's MCP logic. But at the MCP server level, we can make sure our descriptions for get_me
and all tools that depend on get_me
data to work (like PRs) are clear enough so the LLM knows when to call get_me
first.
Possible easy fix:
- Update the tool descriptions for get_me and pull requests explicitly state that owner must come from the cached login returned by get_me – better if get_me is called once per session and reused to save repeat tool calls.
- Should we add a short “requires_auth_user=true” flag (or similar metadata) to all tools that require
get_me
data? Or at the core server logic, check for parameters for indicators thatget_me
will be required? (like a requiredowner
+ it lives in a “write” namespace). Then Agent Mode can detect dependencies automatically.
Affected version
- Remote GitHub MCP Server: 0.5.0 (latest)
- VS Code Stable: 1.101 (latest)
Steps to reproduce the behavior
- Open VS Code stable, connect to the remote GitHub MCP Server, and enter Agent Mode.
- Make some code changes and prompt the agent to open a PR (My PR was merging
feature/add-linter-workflow
to main). - Observe that no
get_me
call appears beforeopen_pull_request
, and the server returns a 404 error.
Might take a few tries to get a similar output. Could try other tool calls that require
get_me
data.
Expected vs actual behavior
Expected – Agent calls get_me
, retrieves the correct username (D1M1TR10S), and successfully creates the pull request with the proper repo owner.
Actual – Agent skips get_me, sends an invalid owner "{my full name}", and the MCP server responds with 404 Not Found.
Logs


