backlog-mcp

An MCP server that gives AI agents structured read/write access to a story-based project backlog. Agents can list stories, read content, update status, and append notes — all backed by plain markdown files that live inside your project repository.

How collaboration works

There is no shared server. The backlog files live in your repo under requirements/, committed and versioned alongside your code. Collaboration between agents, or between an agent and a human, works exactly the way the rest of your codebase does: through git. If two agents update different stories concurrently, git merges them. If they touch the same line, you resolve it like any other merge conflict.

The MCP server is a local process each agent runs for itself. It reads and writes files; git handles the rest.

Install

Homebrew (tap)

brew tap corbym/backlog-mcp
brew install backlog-mcp

You can also install directly without an explicit tap step:

brew install corbym/backlog-mcp/backlog-mcp

Binary release

Download the latest binary for your platform from the Releases page and put it somewhere on your $PATH.

Go install

If you have Go installed:

go install github.com/corbym/backlog-mcp@latest

Build from source

go mod tidy
go build -o backlog-mcp .

Setup

Initialise a requirements/ folder in your project root:

./backlog-mcp init /path/to/your/project/requirements

This creates:

requirements/
  requirements-index.md   # master index — source of truth for epics and story status
  backlog.md              # priority-ordered list of not-done stories
  epic-001-example/
    story-001.md          # example story file

Commit the requirements/ folder to your repo. Edit the files to add your own epics and stories.

Running

./backlog-mcp

The server looks for a requirements/ directory relative to the working directory it is launched from. Claude Code sets the working directory to the project root, so no configuration is needed.

Create a plan file

./backlog-mcp plan [name]

Creates a new plan scaffold in the requirements/ directory. Without a name the file is plan.md; with a name it is plan-<name>.md. If the file already exists a numeric suffix is added (plan-002.md, etc.). Open the file and work with your agent to fill it in before creating stories.

Configuring your MCP client

Prefer a local config file committed to your project root. This scopes the server to the project and means any agent cloning the repo gets the right setup automatically. Only use a global config if you want backlog-mcp available in every project without per-project configuration.

VS Code / GitHub Copilot — add .vscode/mcp.json to your project root:

{
  "servers": {
    "backlog-mcp": {
      "command": "/path/to/backlog-mcp",
      "type": "stdio"
    }
  }
}

Claude Code — add .claude/settings.json to your project root:

{
  "mcpServers": {
    "backlog-mcp": {
      "command": "/path/to/backlog-mcp"
    }
  }
}

For a global fallback (applies to every project), place the same config in ~/.claude/settings.json (Claude Code) or add it to VS Code's user settings.json under the mcp.servers key (GitHub Copilot). Always prefer the local per-project file.

Tools

set_epic_status guards

Setting status to done requires:

1. summary — a completion note, appended as a timestamped entry to the epic file.
2. All stories done — if any stories are still open the call fails and lists them. Pass override_incomplete=true only after the user explicitly confirms the incomplete stories are intentionally omitted.

Moving backwards (e.g. done → in-progress, in-progress → draft) triggers a regression prompt: the agent should offer to create new stories before proceeding. Pass confirm_regression=true only if the user explicitly insists on skipping that step. blocked and deferred are lateral states and can be set freely.

complete_story guards

Acceptance criteria must be set (not the default placeholder) before a story can be completed. Unchecked criteria block completion unless incomplete_items is provided with one explanation per unchecked item. Tick done criteria [x] via set_acceptance_criteria first — do not use incomplete_items to confirm work that is actually finished.

Environment variables

File format

requirements-index.md — one epic section per heading, one story per table row:

## EPIC-001: Combat System — `draft`

| Story | Title | Status | Type |
|-------|-------|--------|------|
| [STORY-001](./epic-001-combat-system/story-001.md) | Basic combat | draft | feature |

backlog.md — priority-ordered numbered list:

1. **STORY-001** — Basic combat
2. **STORY-002** — Enemy AI *(in-progress)*

Story files live at epic-NNN-slug/story-NNN.md under BACKLOG_ROOT.

Story types: feature, bug, chore, spike

Status values: draft, in-progress, done, blocked, deferred

Automated PR backlog agent

A GitHub Actions workflow is included that automatically updates story statuses and appends notes when pull requests are opened or updated. It requires no API keys — only the standard GITHUB_TOKEN.

How it works

On every pull_request event (opened, synchronize) the workflow:

1. Installs the backlog-mcp binary via go install github.com/corbym/backlog-mcp@latest.
2. Scans the PR title and branch name for STORY-NNN IDs.
3. For each matched story, sets status to in-progress (if it was draft and the PR was just opened) and appends a timestamped note with the PR number and title.
4. Commits any changed files under requirements/ back to the PR branch.

Setting it up in your repository

Copy these three files into your repo:

.github/
  actions/
    install/
      action.yml          # composite action — installs the binary via go install
  scripts/
    backlog_agent.py      # deterministic MCP client (Python 3, stdlib only)
  workflows/
    backlog-agent.yml     # the workflow

The files are in the corbym/backlog-mcp repository. No secrets or additional configuration are required beyond a requirements/ folder already being present.

Branch and PR naming

The agent matches stories by STORY-NNN ID. Include the ID in your branch name or PR title:

story-042-short-description          # branch
STORY-042: Short description         # PR title
STORY-042 STORY-043: Short desc      # multiple stories
chore: bump goreleaser to v2         # no story — agent skips cleanly

See CONTRIBUTING.md for the full convention.

Using backlog-mcp with GitHub Copilot agent mode

GitHub Copilot's agent mode in VS Code reads MCP servers from .vscode/mcp.json in your project root. Note the key is "servers", not "mcpServers" (which is the Claude Code convention):

{
  "servers": {
    "backlog-mcp": {
      "command": "/path/to/backlog-mcp",
      "type": "stdio"
    }
  }
}

MCP tools are only available in Agent mode — switch to it via the mode dropdown in Copilot Chat. Once configured, Copilot agent can call list_stories, get_story, add_story_note, and all other backlog tools during a chat session — the same tools the GitHub Actions workflow uses.

Notes

• File writes are atomic (temp file + rename) — a crash mid-write cannot corrupt your files.
• The filesystem is the source of truth. The MCP server never owns the data.