TaskNebula is the open-source, keyboard-first alternative to Jira and Linear. Boards, sprints, and docs that run on your own Postgres — with a built-in MCP server so AI agents work the backlog right beside your team. MIT licensed, self-host with one Docker Compose.
Switch projects, drag issues between stages, and keep details close without fake motion or noisy chrome.
No tool-hopping between a tracker, a planning doc, and an agent console. The work moves through a single surface — humans and agents on the same timeline.
TaskNebula Board
Workspace preview
The flow
Refine launch page hierarchy
Backlog
Plan the sprint in place
To Do
Build alongside agents
MCP tools for agents
Review approved
In Review
Ship and track the outcome
Shipped
Tighten hero, simplify copy, and make the landing feel more premium.
TaskNebula was built for the agent era. Ask questions across your whole workspace, let AI triage and estimate new work, and give external agents real tools through a first-class MCP server.
The /api/ask RAG endpoint answers in natural language over your issues, docs, and comments — grounded in pgvector retrieval, with citations back to the source.
Suggested priority, labels, and effort land on new issues — backed by real OpenAI and Anthropic calls, a cost guard, and tracing. You stay in control: accept or dismiss.
@tasknebula/mcp-server exposes the backlog over MCP — HTTP at /api/mcp or stdio — so external agents create, search, and move issues with revocable keys.
Three containers and one command. Everything lives in your Postgres, exporting is a pg_dump away, and nothing phones home — no vendor lock-in by design.
One Postgres, one web image
Postgres 16 with the pgvector extension powers search; Redis 7 is along for the ride. No Mongo, no Elastic, no sidecar zoo.
Migrations run on boot
The image applies every database migration at startup. Upgrading is pulling a newer tag and restarting — nothing manual.
Health you can probe
GET /api/health returns 200 when the app is ready — wire it to your load balancer, uptime checks, or compose healthcheck.
Pin the version you trust
Set TASKNEBULA_IMAGE=neuraparse/tasknebula:0.14.0 to lock a release, or follow :latest for the bleeding edge.
$ mkdir tasknebula && cd tasknebula$ curl -fsSLo compose.yml https://raw.githubusercontent.com/neuraparse/tasknebula/main/docker-compose.desktop.yml$ docker compose up -dThe web image waits for Postgres, runs migrations, then serves on http://localhost:3000 — open it and finish the first-run admin wizard. Image: neuraparse/tasknebula.
All four are capable tools. Here is exactly where TaskNebula differs — open by default, yours to host, AI-native. Nothing more, nothing less.
Vendor details reflect public documentation and pricing pages as of June 2026 and may change. Jira is a trademark of Atlassian, Linear of Linear Orbit, and Plane of Plane Software — all are solid products, none affiliated with TaskNebula.
No sales gate, no asterisks. If yours is missing, open a GitHub issue — public answers beat private threads.
Yes — the whole app is MIT licensed: boards, sprints, docs, analytics, the AI surfaces, and the MCP server. No open-core split, no paywalled tier, no per-seat fee. Self-host it and run it forever for unlimited users; fork it if you want to.
That is the default. Bring Docker and you get three containers — Postgres 16 with pgvector, Redis 7, and the neuraparse/tasknebula web image. Migrations run on boot, GET /api/health is your probe endpoint, and a one-line compose quickstart gets you to localhost:3000. There is no hosted-only feature you lose by running it yourself.
AI is optional — every board, sprint, and doc works with no model keys configured. When you want it, you bring your own OpenAI or Anthropic key as environment variables, so spend and data residency stay yours. There is a cost guard so a runaway prompt cannot run up your bill.
We build them in the open. The data model and endpoints — triage suggestions, pgvector semantic search, and the /api/ask RAG endpoint — are live, and they make real OpenAI and Anthropic calls. Some surfaces are still labeled preview while we harden them; we would rather say that here than have you discover it after install.
An MCP (Model Context Protocol) server exposes TaskNebula as tools — create, search, move, and comment on issues — so agents like Claude Code and Cursor can drive your tracker. It ships with the repo, HTTP at /api/mcp or stdio via @tasknebula/mcp-server, and is MIT licensed. Heads up: end-to-end agent auth depends on the REST API accepting API keys, which is still being wired, so treat it as preview today.
Yes — there are importers for Jira, Linear, GitHub, and CSV. Each runs with a preview step so you can inspect the field mapping before anything is written, and you can keep your old tool read-only while you cut over.
The core tracker — issues, boards, sprints, docs, search, permissions, REST API, and OpenAPI spec — is solid and self-hostable. Some edges are still maturing: OAuth login needs a DB adapter, so use credentials auth for now; realtime collab env vars are not yet plumbed into the Docker images; and parts of the AI/MCP surface are preview. We track all of it openly in docs/STATUS.md.
You decide. Self-hosted, everything lives in your Postgres on your infrastructure and exports with a plain pg_dump — no telemetry phones home. Inside the product, workspaces are isolated per organization with role-based access, permission schemes, and audit logs.
Still unsure? Ask on GitHub or read the live status in the docs.
Create a workspace in minutes, or run the whole thing on your own box tonight — either way, the code is yours to read.