Skip to main content
Glama

backlog-mcp

Ein Aufgaben-Backlog-MCP-Server für LLM-Agenten. Funktioniert mit jedem MCP-Client — Claude, Kiro, Cursor, Codex usw.

Agenten erstellen Aufgaben, verfolgen den Fortschritt, hängen Artefakte an und durchsuchen alles. Menschen erhalten einen Echtzeit-Web-Viewer, um zu sehen, was Agenten tun.

Läuft sofort lokal. Kann auch auf Cloudflare Workers + D1 selbst gehostet werden, für ein kostenloses, ständig verfügbares Remote-Backlog, das von jedem Gerät oder MCP-Client aus zugänglich ist.

Schnellstart: Sagen Sie Ihrem LLM: Add backlog-mcp to .mcp.json and use it to track tasks

Live-Demo: backlog-mcp-viewer.pages.dev — die Viewer-UI, verbunden mit einer echten gehosteten Instanz

backlog-mcp web viewer

Was ist enthalten

Dies ist ein Monorepo mit 3 Paketen:

Paket

npm

Was es tut

packages/server

backlog-mcp

MCP-Server, HTTP-API, CLI

packages/viewer

—

Web-UI, basierend auf @nisli/core

packages/shared

—

Gemeinsame Entitätstypen und ID-Dienstprogramme

Der Viewer wurde mit Nisli erstellt, einem reaktiven Web-Component-Framework ohne Abhängigkeiten, das als @nisli/core veröffentlicht wurde. Nisli begann in diesem Repository und existiert nun separat.

Related MCP server: Task Manager MCP Server

Installation

Fügen Sie es zu Ihrer MCP-Konfiguration hinzu (.mcp.json oder Ihre MCP-Client-Konfiguration):

{
  "mcpServers": {
    "backlog": {
      "command": "npx",
      "args": ["-y", "backlog-mcp"]
    }
  }
}

Self-Hosting auf Cloudflare (Optional)

Hosten Sie Ihr eigenes, ständig verfügbares Remote-Backlog kostenlos mit Cloudflare Workers + D1. Von jedem Gerät oder MCP-Client aus zugänglich — kein lokaler Serverprozess erforderlich.

Voraussetzungen: Cloudflare-Konto (kostenloser Tarif reicht aus), wrangler CLI, eine GitHub OAuth App.

# 1. Clone and install
git clone https://github.com/gkoreli/backlog-mcp.git
cd backlog-mcp
pnpm install

# 2. Create the D1 database
cd packages/server
npx wrangler d1 create backlog
# Copy the database_id into packages/server/wrangler.jsonc

# 3. Apply the D1 migrations
npx wrangler d1 execute backlog --remote --file=migrations/0001_initial.sql
npx wrangler d1 execute backlog --remote --file=migrations/0002_oauth_refresh_tokens.sql

# 4. Set secrets
npx wrangler secret put JWT_SECRET          # any strong random string
npx wrangler secret put API_KEY             # your personal API key (for Claude Desktop)
npx wrangler secret put GITHUB_CLIENT_ID    # GitHub OAuth App client ID
npx wrangler secret put GITHUB_CLIENT_SECRET
npx wrangler secret put ALLOWED_GITHUB_USERNAMES  # comma-separated: "you,youralt"

# 5. Deploy
npx wrangler deploy

GitHub OAuth App-Einrichtung: Gehen Sie zu GitHub → Settings → Developer settings → OAuth Apps → New. Setzen Sie die Callback-URL auf https://<your-worker>.workers.dev/oauth/github/callback.

Nach der Bereitstellung fügen Sie es zu Ihrer MCP-Client-Konfiguration hinzu:

{
  "mcpServers": {
    "backlog": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://<your-worker>.workers.dev/mcp"]
    }
  }
}

Claude.ai und ChatGPT können sich direkt über die Remote-MCP-URL verbinden — kein lokaler Prozess erforderlich. GitHub OAuth-Sitzungen verwenden rotierende Refresh-Tokens, sodass Clients den Zugriff erneuern können, ohne sich täglich neu authentifizieren zu müssen.


Web-Viewer

Öffnen Sie http://localhost:3030 — immer verfügbar, wenn der Server läuft.

Funktionen:

  • Layout mit geteiltem Bereich für Aufgabenliste und Detailansicht

  • Spotlight-Suche mit hybrider Text- + semantischer Übereinstimmung

  • Echtzeit-Updates via SSE

  • Aktivitäts-Timeline

  • Filtern nach Status, Typ, Epic

  • Markdown-Rendering im GitHub-Stil mit Mermaid-Diagrammen

  • URL-Zustandspersistenz

Die Viewer-UI wurde mit Nisli (@nisli/core) erstellt.

Entitätstypen

5 Entitätstypen, alle als Markdown-Dateien mit YAML-Frontmatter gespeichert:

Typ

Präfix

Zweck

Task

TASK-0001

Arbeitselemente

Epic

EPIC-0001

Gruppen zusammengehöriger Aufgaben

Folder

FLDR-0001

Organisationscontainer

Artifact

ARTF-0001

Angehängte Ausgaben (Recherche, Designs, Protokolle)

Milestone

MLST-0001

Zeitgebundene Ziele mit Fälligkeitsdaten

Statuswerte: open, in_progress, blocked, done, cancelled

Beispiel für eine Aufgabendatei:

---
id: TASK-0001
title: Fix authentication flow
status: open
epic_id: EPIC-0002
parent_id: FLDR-0001
references:
  - url: https://github.com/org/repo/issues/123
    title: Related issue
evidence:
  - Fixed in PR #45
---

The authentication flow has an issue where...

MCP-Tools

backlog_list

backlog_list                              # Active tasks (open, in_progress, blocked)
backlog_list status=["done"]              # Completed tasks
backlog_list type="epic"                  # Only epics
backlog_list epic_id="EPIC-0002"          # Tasks in an epic
backlog_list parent_id="FLDR-0001"        # Items in a folder
backlog_list query="authentication"       # Search across all fields
backlog_list counts=true                  # Include counts by status/type
backlog_list limit=50                     # Limit results

backlog_get

backlog_get id="TASK-0001"                # Single item
backlog_get id=["TASK-0001","EPIC-0002"]  # Batch get

backlog_create

backlog_create title="Fix bug"
backlog_create title="Fix bug" description="Details..." epic_id="EPIC-0002"
backlog_create title="Q1 Goals" type="epic"
backlog_create title="Research notes" type="artifact" parent_id="TASK-0001"
backlog_create title="v2.0 Release" type="milestone" due_date="2026-03-01"
backlog_create title="Fix bug" source_path="/path/to/spec.md"  # Read description from file

backlog_update

backlog_update id="TASK-0001" status="done"
backlog_update id="TASK-0001" status="blocked" blocked_reason=["Waiting on API"]
backlog_update id="TASK-0001" evidence=["Fixed in PR #45"]
backlog_update id="TASK-0001" parent_id="FLDR-0001"
backlog_update id="MLST-0001" due_date="2026-04-01"

backlog_delete

backlog_delete id="TASK-0001"             # Permanent delete

backlog_search

Hybride Volltext- + semantische Suche mit Relevanzbewertung:

backlog_search query="authentication bug"
backlog_search query="design decisions" types=["artifact"]
backlog_search query="blocked tasks" status=["blocked"] limit=10
backlog_search query="framework" sort="recent"
backlog_search query="search ranking" include_content=true

backlog_context

Erhalten Sie umfassenden Kontext für eine Aufgabe — übergeordnetes Epic, Geschwister, Kinder, Querverweise, Rückverweise, kürzliche Aktivitäten und semantisch verwandte Elemente:

backlog_context task_id="TASK-0001"
backlog_context task_id="TASK-0001" depth=2          # Grandparent/grandchildren
backlog_context query="search ranking improvements"   # Find by description
backlog_context task_id="TASK-0001" include_related=false  # Skip semantic search

write_resource

Bearbeiten Sie vorhandene Dateien auf dem MCP-Server. Alle Erstellungen erfolgen über backlog_create.

# Edit task body (use str_replace — protects frontmatter)
write_resource uri="mcp://backlog/tasks/TASK-0001.md" \
  operation={type: "str_replace", old_str: "old text", new_str: "new text"}

# Insert after a specific line
write_resource uri="mcp://backlog/tasks/TASK-0001.md" \
  operation={type: "insert", insert_line: 5, new_str: "inserted line"}

# Append to a file
write_resource uri="mcp://backlog/resources/log.md" \
  operation={type: "append", new_str: "New entry"}

Operationen: str_replace (exakte Übereinstimmung, muss eindeutig sein), insert (nach Zeilennummer), append (Dateiende).

Funktionsweise

Das Ausführen von npx -y backlog-mcp (die Standard-MCP-Konfiguration) bewirkt Folgendes:

  1. Startet einen persistenten HTTP-Server als losgelösten Hintergrundprozess — bedient sowohl den MCP-Endpunkt (/mcp) als auch den Web-Viewer (/) auf Port 3030

  2. Überbrückt stdio dorthin — Ihr MCP-Client kommuniziert über stdio, das via mcp-remote an den HTTP-Server weitergeleitet wird

  3. Auto-Updates: npx -y zieht immer die neueste veröffentlichte Version. Wenn der laufende Server eine ältere Version ist, wird er automatisch heruntergefahren und mit der neuen neu gestartet

  4. Resiliente Wiederherstellung: Wenn die Brücke die Verbindung verliert, startet ein Supervisor sie mit exponentiellem Backoff neu (bis zu 10 Wiederholungsversuche). Verbindungsfehler wie ECONNREFUSED werden erkannt und automatisch behandelt

Der HTTP-Server bleibt über Agentensitzungen hinweg bestehen — mehrere MCP-Clients können ihn gemeinsam nutzen. Der Web-Viewer ist immer unter http://localhost:3030 verfügbar.

CLI

Alle Befehle via npx:

npx backlog-mcp                # Start stdio bridge + auto-spawn HTTP server (default)
npx backlog-mcp status         # Check server status
npx backlog-mcp stop           # Stop the server
npx backlog-mcp version        # Show version
npx backlog-mcp serve          # Run HTTP server in foreground (optional, see below)

Beispielausgaben:

$ npx backlog-mcp status
Server is running on port 3030
Version: 0.44.0
Data directory: /Users/you/.backlog
Task count: 451
Uptime: 3515s
Viewer: http://localhost:3030/
MCP endpoint: http://localhost:3030/mcp

$ npx backlog-mcp stop
Stopping server on port 3030...
Server stopped

$ npx backlog-mcp status
Server is not running

Die CLI existiert für Menschen, um den Hintergrundserver, den Agenten verwenden, zu inspizieren und zu verwalten. Da der Standardmodus einen losgelösten Prozess startet, benötigen Sie status, um ihn zu überprüfen, und stop, um ihn herunterzufahren.

serve führt den HTTP-Server im Vordergrund statt losgelöst aus — nützlich für Debugging, Docker-Container oder den Betrieb ohne MCP-Client. Bei normaler Verwendung benötigen Sie dies nie; der Standardbefehl erledigt alles.

Konfiguration

BACKLOG_DATA_DIR=~/.backlog    # Where to store tasks (default: data/tasks/)
BACKLOG_VIEWER_PORT=3030       # HTTP server port

Erstellen Sie eine .env-Datei für die lokale Entwicklung — siehe .env.example.

Entwicklung

git clone https://github.com/gkoreli/backlog-mcp.git
cd backlog-mcp
pnpm install
pnpm build          # Build all packages
pnpm test           # Run all workspace tests
pnpm dev            # Server + viewer with hot reload

Architektur

packages/
├── server/       # MCP server, search, context hydration, storage
├── viewer/       # Web UI built with @nisli/core
└── shared/       # Entity types, ID utilities
docs/
├── adr/              # backlog-mcp architecture decision records
└── framework-adr/    # Pointer to Nisli ADRs

Backlog-ADRs dokumentieren wichtige Designentscheidungen. Siehe docs/adr/README.md für den vollständigen Index. Nisli-ADRs befinden sich im Nisli-Repository.

Lizenz

MIT

Available Tools

23 tools
backlog_accept_adrA

Use when ratifying an existing proposed ADR.

Inputs: id (required), as.

backlog_attach_artifactB

Use when attaching an artifact to a project item. parent_id is required.

Inputs: title (required), parent_id (required), as, path, +3 more.

backlog_block_taskC

Use when blocking a task with a reason.

Inputs: id (required), blocked_reason (required), as.

backlog_capture_promptC

Use when preserving a verbatim human directive as a chronological project prompt.

Inputs: title (required), content (required), as, date, +3 more.

backlog_capture_referenceA

Use when recording an external technology, prior-art system, or benchmark this project builds on — with authoritative citations and the project's adopt/park/reject relationship to it.

Inputs: title (required), content (required), bears_on (required), category (required), +4 more.

backlog_capture_requirementC

Use when recording a human or system requirement in the current project.

Inputs: title (required), content (required), as, date, +11 more.

backlog_complete_taskC

Use when completing a task, optionally with evidence.

Inputs: id (required), as, evidence.

backlog_consolidation_candidatesA

List clusters of episodic memories that are ripe for consolidation into durable knowledge. Consolidator workflow: (1) call this and take ripe bundles; (2) per bundle, read members (backlog_get on MEMO- ids for depth), then write ONE narrative memory via backlog_remember({ layer: "semantic"|"procedural", derived: true, entity_refs: [member MEMO- ids + key source entities], context }) — a self-contained story, not fragments; (3) retire the members via backlog_forget({ ids }) so they stop appearing here while staying auditable; (4) track your progress with a backlog_remember state_key like "consolidation.watermark.".

Inputs: home, limit, context, min_count, +4 more.

backlog_contradictionsA

List structural contradictions in memory: sets of ≥2 LIVE memories that share one state_key (e.g. "db.primary"), which should never happen — a new memory with a state_key auto-expires the previous holder. Each set means two beliefs about the same fact are both active. Resolution is yours, never automatic: pick the correct member, then either backlog_remember({ content, state_key, supersedes: }) to record the right value and retire the rest, or backlog_forget({ ids: [] }) to expire the wrong ones. Read members with backlog_get for full context before deciding. Set candidates: true to list semantic collision candidates instead: nearby live facts that deserve review, never contradiction verdicts.

Inputs: home, candidates, project_root.

backlog_create_workB

Use when creating a project work item. Pass parent_id when known; parentless work surfaces as unfiled at wakeup.

Inputs: title (required), as, content, parent_id, +1 more.

backlog_deleteC

Delete an item permanently.

Inputs: id (required), as, home, project_root.

backlog_forgetA

Retract memories — soft-expire them so recall stops returning them (the record stays auditable in the viewer). Use when knowledge is wrong or obsolete and there is no replacement (if there IS a replacement, use backlog_remember with supersedes instead). expired:true garbage-collects already-expired memories.

Inputs: ids, home, layer, context, +3 more.

backlog_getA

Get full details by ID. Accepts task IDs (TASK-0001, EPIC-0002), document paths (README.md, docs/adr/0001-example.md), or MCP resource URIs (mcp://backlog/resources/design.md). Works for any item regardless of status. Pass context:true when starting work on an entity to also see its relational neighborhood as stubs.

Inputs: id (required), home, depth, context, +1 more.

backlog_organize_folderA

Use when creating a folder to organize project items. Pass parent_id when known; parentless work surfaces as unfiled at wakeup.

Inputs: title (required), as, content, parent_id.

backlog_pause_cronB

Use when pausing a scheduled intake.

Inputs: id (required), as.

backlog_recallA

Recall memories — knowledge and episodes captured across sessions. Returns STUBS (title + one-line digest + provenance) by default; expand interesting ones with backlog_get(MEMO-id), or pass full:true for bodies. Weigh a stub's trust BEFORE hydrating: age_days (on the knowledge's own timeline), uses/idle_days (recall demand), supersedes (this is a correction), derived (consolidator inference), kind (current/historical/plan/preference/timeless). Old + never-used = treat as hypothesis, not truth. Distinct from backlog_search (live entities). Use to answer "how do we deploy?", "have I hit this before?", "what did I finish about X?". Memories point back to source entities via entity_id.

Inputs: query (required), full, home, tags, +5 more.

backlog_rememberA

Write a durable memory — a stable fact, a procedure, or a preference you should know next session. Use when you learn something worth keeping: "this repo deploys via wrangler", "Goga prefers terse evidence bullets". To CORRECT existing knowledge, pass supersedes (the old MEMO- id is expired, lineage kept) or state_key (previous holders of the same evolving fact are closed). The optional collision_candidates receipt is tri-state: [] means the advisory scan completed clean; a non-empty array means nearby facts deserve adjudication; absent means the advisory scan did not run or failed, not that the write failed. Do not use for task events — completions are captured automatically.

Inputs: title (required), content (required), as, home, +11 more.

backlog_resume_cronC

Use when resuming a scheduled intake.

Inputs: id (required), as.

backlog_schedule_cronB

Use when scheduling recurring project intake. Pass parent_id when known; parentless work surfaces as unfiled at wakeup.

Inputs: title (required), command (required), schedule (required), as, +3 more.

backlog_supersede_adrA

Use when a newer ADR replaces an accepted or living ADR while preserving lineage.

Inputs: superseded_id (required), replacement_id (required), as.

backlog_wakeupA

Dense session-start briefing: active tasks, current epics, project constraints (requirements as stubs, violated/at-risk first — treat these as standing product intent), recent completions (with evidence snippets), recent activity, and the home-wide unfiled work count. No focal entity required — use this at the start of every session to understand what you were working on. Optional scope narrows the entity sections to a folder, milestone, or epic; unfiled remains home-wide because parentless work has no subtree ancestry. Optional operation resumes a mid-flight operation document: its live state becomes the briefing's FOCUS centerpiece and the rest of the briefing yields budget to it.

Inputs: home, scope, operation, max_activity, +4 more.

write_resourceA

Use when you want backlog-mcp to validate and canonically persist an existing entity edit before reporting success. For ordinary repository prose edits, use your native Edit tool; reconciliation updates indexes and diagnostics afterward. Create and transition entities through the substrate-declared intent tools.

  • The append command will add content to the end of the body, automatically adding a newline if needed. Notes for using the str_replace command:

  • The old_str parameter should match EXACTLY one or more consecutive lines from the original body. Be mindful of whitespaces!

  • If the old_str parameter is not unique in the body, the replacement will not be performed. Include enough context to make it unique.

  • The new_str parameter should contain the edited lines that should replace the old_str

Inputs: id (required), operation (required), as, home, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.2/5.0

    Scored across 23 tools

    Disambiguation3/5

    Some tools are clearly distinct (backlog_get, backlog_delete, backlog_search), but overlap exists among 'capture' tools (backlog_capture_prompt, backlog_capture_reference, backlog_capture_requirement) and 'create' tools (backlog_create_work, backlog_organize_folder, backlog_schedule_cron) that may confuse selection. Also, backlog_wakeup and backlog_recall both serve retrieval purposes, but their contexts differ; the descriptions do help to differentiate these.

    Naming Consistency3/5

    All tool names are lowercase with underscores, but the verb style varies: some use verb_noun (backlog_get, backlog_delete), others use verb_object (backlog_complete_task, backlog_attach_artifact), and some lack a clear verb-object structure (backlog_consolidation_candidates, backlog_contradictions). The consistent prefix 'backlog_' helps, but patterns like 'create' vs 'capture' and 'organize' are not uniform.

    Tool Count4/5

    23 tools is a fairly large set, but it covers the domain of a project management/knowledge memory server well. Each tool serves a distinct function — CRUD for work items, ADR management, memory operations, scheduling, and organization. Slightly heavy, but justified given the breadth of features (tasks, ADRs, memories, scheduling).

    Completeness4/5

    The tool surface is extensive, covering work item lifecycle (create, get, delete, complete, block), ADR handling (accept, supersede), memory management (remember, recall, forget, consolidation), and project organization (folders, attachments, requirements). Minor gaps include no explicit 'update_work' or 'unblock_task', but these may be handled via write_resource or other mechanisms.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers