Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@io.github.RohitYajee8076/backburner run npm test in background"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Put your AI agent's slow work on the back burner. Keep cooking.
Background tasks for AI agents that outlive the conversation — start a long job, close the client, and the result is still waiting when you come back.
Durable & Restart-Proof ◦ Zero Infrastructure ◦ MCP Tasks (2026-07-28) ◦ Windows & Unix
📦 PyPI • 🗂️ MCP Registry • 🐛 Issues • 📄 MIT
v1.0 — implements the official MCP Tasks extension
(SEP-2663,
io.modelcontextprotocol/tasks). A Tasks-capable client can turn a
start_task call into a durable task and drive it with tasks/get,
tasks/update, and tasks/cancel — the standard async-job protocol — while
the five plain tools keep working for every other client. Built against the
2026-07-28 spec (mcp 2.0).
v0.2.1 — output with non-ASCII characters (✓, emoji, any non-English text) no longer crashes tasks on Windows.
v0.2.0 — exit_code is no longer reported for cancelled/timed-out tasks
(it was an artifact of the kill, not a real result); new animated demo below.
v0.1.x — first release: 5 tools, task timeouts, command allow/deny policy.
Listed on the official MCP Registry as io.github.RohitYajee8076/backburner.
backburner is an MCP server that gives any AI assistant — Claude, ChatGPT,
Gemini, GitHub Copilot, Cursor, and any other MCP client — the ability to run
long shell commands as background tasks — start a test suite, a build, a
scrape, a batch job — then keep working and check back for the results, instead
of sitting frozen until it finishes.
Related MCP server: Background Process MCP
Because that lives inside the conversation — it disappears the moment the session ends. Close the chat, restart the client, reboot the laptop, and any in-session background work (and its output) is gone.
backburner keeps every task and its full output on disk (SQLite +
per-task log files under ~/.backburner/), so your work outlives the session
that started it:
Start now, collect later — even in a different chat. A task you launch today is still listed, with its result, in a brand-new session tomorrow.
Restart-proof. State survives the server, the client, and the machine
restarting. Finished tasks keep their output; a task cut off by a crash is
honestly marked interrupted, never silently dropped.
No waiting, no blocking. A 10-minute tool call no longer freezes the conversation or times out and loses the work.
See it for yourself — a real two-process proof (no mock-ups):
python docs/demo_restart.pyIt starts a job in one process, exits, then a separate process — which never saw the task id — finds the finished work waiting on disk.
Built on the MCP Tasks pattern, formalized in the 2026-07-28 spec release
(SEP-2663):
backburner speaks it natively (tasks/get / tasks/update / tasks/cancel)
and exposes the same engine as plain tools, so it works with every client
today.
Tool | What it does |
| Run a shell command in the background, returns a task id immediately |
|
|
| Captured output — works mid-run too, so you can peek at progress |
| Kill the task and its whole process tree |
| Recent tasks, newest first |
Survives restarts — tasks are tracked in SQLite under ~/.backburner/;
output is captured to per-task log files. If the server dies mid-task,
orphaned tasks are honestly marked interrupted, never silently lost.
Real cancellation — kills the full process tree (worker processes included), on Windows and Unix.
Peek at live progress — task_result on a running task returns the
output so far.
Timeouts — pass timeout_seconds and a runaway task is killed and
honestly marked timed_out instead of hanging forever.
Command policy — restrict what the AI may run with environment variables (regexes, comma-separated; deny always wins):
BACKBURNER_ALLOW="^pytest,^npm (test|run build)" # only these may run
BACKBURNER_DENY="rm -rf,shutdown,format" # these never runZero infrastructure — stdlib only (SQLite, subprocess, threads). No Redis, no Celery, no Docker.
Tested — a pytest suite covers the full job lifecycle: completion, failure, cancellation, timeouts, crash recovery, and the command policy.
backburner is a standard stdio MCP server — it works with any MCP-compatible
client, including:
Claude Code · Claude Desktop · OpenAI (ChatGPT desktop / Agents SDK) · Google Gemini (Gemini CLI) · GitHub Copilot (VS Code) · Cursor · Windsurf · Cline · Zed — and any other client that speaks MCP.
First install the package:
pip install backburner-mcpclaude mcp add backburner -- python -m backburner.serverMost clients use the same standard config block — add backburner to your
client's MCP config (see your client's docs for where that file lives):
{
"mcpServers": {
"backburner": {
"command": "python",
"args": ["-m", "backburner.server"]
}
}
}backburner executes the shell commands the AI sends it, with your user's
permissions. That is its job — but treat it like giving your agent a
terminal. Run it only with clients whose tool-use you review/approve,
prefer permission modes that require confirmation for start_task, and
use BACKBURNER_ALLOW / BACKBURNER_DENY to scope what may run.
Task timeouts and max-runtime limits
Allowlist/denylist for commands
PyPI release — pip install backburner-mcp
Listed on the official MCP Registry
MCP Tasks extension (spec 2026-07-28, SEP-2663) — native tasks/get /
tasks/update / tasks/cancel alongside the plain tools
Task push updates (notifications/tasks) — live status without polling
Local web dashboard — watch tasks live in the browser
Structured progress reporting (parse % / step markers from output)
MIT
Cancel a running task, killing its whole process tree.
Inputs: task_id (required).
List recent tasks, newest first.
Inputs: limit. Returns: result
Start a shell command as a background task and return immediately.
Args: command: The shell command to run (e.g. "pytest -q" or "npm run build"). cwd: Working directory for the command. Defaults to the server's cwd. timeout_seconds: If set, the task is killed and marked 'timed_out' when it runs longer than this. Recommended for unattended jobs.
Returns the new task's id and initial status. The command keeps running after this call returns — use task_status / task_result to follow it.
Inputs: command (required), cwd, timeout_seconds.
Get a task's captured output (stdout+stderr merged).
Args: task_id: The task to inspect. Works for finished AND still-running tasks, so you can peek at live progress. tail_lines: How many trailing lines of output to return.
Inputs: task_id (required), tail_lines.
Check on a task: working, completed, failed, cancelled, timed_out, or interrupted.
Inputs: task_id (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 5 tools
Each tool targets a distinct lifecycle stage: start_task initiates, task_status reports state, task_result retrieves output, cancel_task terminates, and list_tasks enumerates. No two tools overlap in purpose.
All tool names follow a consistent verb_noun pattern in snake_case: start_task, task_status, task_result, cancel_task, list_tasks. There are no deviations or mixed conventions.
Five tools is a perfect scope for a background task manager: create, monitor, retrieve output, cancel, and list. Each tool has a clear role, and no redundant or missing entries.
The tool set covers the full lifecycle of a background task: start, poll status, fetch output, cancel, and list. The timeout behavior in start_task also handles timed-out tasks. No obvious gaps.
Durable background job execution, async task scheduling, and state persistence for AI agents.
Operate Linux, macOS and Windows from your LLM. Every action runs through an auditable allowlist.
Remote shell and detached long-running jobs on your own machines — no SSH, open ports or VPN.
Persistent work tracking for AI agents: tasks, status and history that follow you across machines
