backlog

An MCP server that wraps GitHub Issues as a lightweight backlog for Claude Code. Lets Claude create, track, and resolve work items using your GitHub repo as the backend — no extra database or token management required.

How it works

The server exposes 12 tools to Claude via the Model Context Protocol:

All GitHub operations go through the gh CLI — no personal access tokens to manage.

Prerequisites

• Python 3.10+
• uv installed (curl -LsSf https://astral.sh/uv/install.sh | sh)
• GitHub CLI installed and authenticated (gh auth login)

Installation

uvx gh-backlog-mcp --setup

That's it. uvx fetches the package from PyPI on demand — no cloning, no pip install. The setup command:

1. Verifies gh is installed and authenticated
2. Registers the MCP server in ~/.claude.json
3. Prints a CLAUDE.md snippet to add to your project

Labels are created automatically on first use in each repo. Restart Claude Code after setup to pick up the server.

Configuration

Setup registers the server in ~/.claude.json using uvx:

{
  "mcpServers": {
    "backlog": {
      "type": "stdio",
      "command": "uvx",
      "args": ["gh-backlog-mcp"]
    }
  }
}

The repo is auto-detected from the current working directory via gh repo view. The server targets whichever GitHub repo you have open — no configuration needed.

Labels

The server auto-creates these label groups on first use in each repo:

Branch Scoping

Issues are automatically scoped to the current git branch via branch:X labels. When you create or list backlog items, the server detects your current branch and applies the corresponding label.

• Auto-detect — create_backlog_item, list_backlog_items, search_backlog, and create_spike all default to the current git branch
• Cross-branch view — pass branch="all" to any of these tools to see/create items across all branches
• Explicit branch — pass branch="feature/foo" to target a specific branch regardless of what you have checked out
• Graceful degradation — if branch detection fails (not in a git repo, detached HEAD), scoping is silently skipped

After merging a branch

Use migrate_branch_issues to transfer open issues from a merged feature branch to the target branch:

migrate_branch_issues(from_branch="feature/auth-rewrite")

The to_branch defaults to the current git branch, so after merging feature/auth-rewrite into main, just call this from main and the issues will be relabeled automatically.

Research Spikes

A research spike is a formal investigation item that signals to the agent picking it up: enter plan mode, do web research, and produce a written output before any implementation begins.

Creating a spike

create_spike(
  title="Evaluate vector DB options for embedding search",
  goal="Determine which vector database best fits our latency and cost requirements",
  research_questions="- What are the top open-source options?\n- How do they compare on query latency at 1M vectors?\n- What are the operational costs?",
  success_criteria="Output doc must include a comparison table and a recommendation with rationale",
  priority="high"
)

The issue is created with both research and spike labels. The body includes a structured template (Goal, Background, Research Questions, Success Criteria, Output path).

Agent workflow for spikes

When an agent calls start_working_on on a spike, it receives explicit instructions to:

1. Enter plan mode
2. Call check_research to find any prior work on the topic
3. Use WebSearch/WebFetch to answer the research questions
4. Call save_research_output with the synthesized findings
5. Call complete_backlog_item to close the issue

Research output files

save_research_output writes findings to research/<issue-number>-<slug>.md in the current working repository (determined via git rev-parse --show-toplevel). It also:

• Creates the research/ directory if it doesn't exist
• Adds research/ to .gitignore automatically so outputs are never accidentally committed
• Posts a comment on the GitHub issue linking to the file

Checking prior research

check_research(topic="vector database")

Searches two sources:

1. Local research/*.md files — keyword match against filename and first 500 chars
2. GitHub spike issues (open and closed) matching the topic

Use this before creating a new spike or before implementing something that may have been researched previously.