LastPing MCP

Мониторинг с помощью AI-агента

Документация

Most monitoring watches a thing and tells you when it looks wrong. LastPing waits for a thing to check in and tells you when it doesn't. That inversion is the whole product: a job that breaks can't send you an error, but it can fail to send you anything — and absence is the one signal a broken process can still produce.

This repository holds the open-source pieces: the lastping CLI and the MCP server. The hosted service they talk to is at lastping.dev, free for individuals.

Install

curl -fsSL https://raw.githubusercontent.com/tp322d/lastping-app/main/install.sh | sh

No Go toolchain needed — that pulls a prebuilt binary for macOS and Linux, on amd64 and arm64, and verifies its checksum. Windows builds are on the releases page.

If you do have Go:

go install github.com/tp322d/lastping-app/cmd/lastping@latest

`lastping run` — reporting you can't forget

Put it in front of whatever you already run:

lastping run --monitor <monitor-id> -- python nightly_etl.py
lastping run --monitor <monitor-id> -- ./backup.sh
lastping run --monitor <monitor-id> -- claude

It sends a start ping, runs your command untouched, and reports the exit code when it finishes — success on 0, failure on anything else, with the tail of stderr attached so the alert says why.

Three properties worth knowing, because they are the difference between a monitoring wrapper you can trust in production and one you remove after a bad night:

Your exit code always propagates. The wrapper exits with whatever your command exited with, so CI behaves exactly as it did before you added it.
A failed ping never touches your command. If LastPing is unreachable, your job still runs, still writes its output, still exits normally.
Interactive stays interactive. stdin and stdout are handed over as file descriptors, so wrapping a REPL or an agent session works.

Why a wrapper rather than an instruction? Because anything advisory decays. An AI agent told to report on every task will stop doing it, and a cron line you meant to add a curl to never gets it. A wrapper reports from the process lifecycle, so nothing depends on anybody remembering.

Traces (ships with the next server release)

lastping run always configures your wrapped command's OpenTelemetry exporter, in its environment only, so an auto-instrumented agent can export its own trace spans with no code change:

OTEL_EXPORTER_OTLP_TRACES_ENDPOINT — the header-free monitor-URL form (<ping url>/v1/traces) when LASTPING_API_KEY is not set, so a headerless exporter can still authenticate; the ping host's /v1/traces (the Bearer form) when it is set.
OTEL_RESOURCE_ATTRIBUTES — lastping.monitor_id=<id>,lastping.run_id=<rid> appended to whatever you already set, so a trace's spans join the same run the surrounding pings report.
OTEL_EXPORTER_OTLP_HEADERS — Authorization=Bearer <your key>, only when LASTPING_API_KEY is set and you have not already set that variable yourself.

Any of the three you already set is left alone. The key is never used to authenticate a ping; the ping URL stays unauthenticated by design, as above.

MCP server — let an agent set up its own monitoring

Claude Desktop's config file runs local programs only, so LastPing connects through the mcp-remote bridge. Add this to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "lastping": {
      "command": "/ABSOLUTE/PATH/TO/npx",
      "args": [
        "-y", "mcp-remote",
        "https://mcp.lastping.dev/mcp",
        "--header", "Authorization:${LASTPING_AUTH}"
      ],
      "env": {
        "PATH": "/FOLDER/THAT/HOLDS/npx:/usr/bin:/bin",
        "LASTPING_AUTH": "Bearer lp_your_key"
      }
    }
  }
}

Needs Node.js. Claude Desktop does not read your shell's PATH, so give it the full path: run which npx in a terminal and put the result in command (with nvm it looks like /Users/you/.nvm/versions/node/v22.11.0/bin/npx), and its folder at the front of PATH in env.
Quit and reopen Claude Desktop; it reads the file only at start-up. Settings, Connectors then lists lastping.
Adding LastPing as a custom connector (Settings, Connectors, Add custom connector) needs a sign-in LastPing does not offer yet, so the bridge above is the way for now. Cowork uses the same Desktop entry.

Claude Code connects directly, no bridge: claude mcp add --transport http --scope user lastping https://mcp.lastping.dev/mcp --header "Authorization: Bearer <key>". Cursor, Windsurf, Codex CLI, Gemini CLI and other clients: lastping.dev/mcp/#connect.

The hosted server is the recommended path, and it always carries the current tool set. Claude Code, Cursor, Windsurf, Codex CLI and Gemini CLI connect to it with a URL and an API key, with nothing to install. Claude Desktop needs one extra step: the mcp-remote bridge above, which needs Node.js.

A stdio binary is also here if you would rather run it yourself:

go install github.com/tp322d/lastping-app/cmd/lastping-mcp@latest

Tools in this repository's stdio binary (50)

Monitors: create_monitor · get_monitor · list_monitors · update_monitor · delete_monitor · pause_monitor · resume_monitor · snooze_monitor

Discovery: discover_monitors_reconcile

Reporting: get_ping_instructions · declare_run_expectations

Incidents & runs: list_incidents · get_run_history · get_run (one run's full timeline, assertion verdicts and spans) · list_runs (runs across every monitor, traced runs included, with filters) · get_incident (one incident's recorded timeline)

Tracing: get_trace_setup (the set-up steps for one tool, from the server) · create_ingest_key (a tracing key bound to one monitor; a write key is enough) · get_trace_diagnostics (why a sent span was refused)

Agent observability: get_agent_dependencies · get_agent_usage · list_dependencies · list_discovered_agents · adopt_discovered_agent

The failure loop: list_open_incidents · add_incident_note

Alert routing: set_route · delete_route

Delivery log: list_deliveries (ships with the next server release) — recent alert deliveries across every monitor, no paging

Destinations: list_destinations · create_destination · update_destination · test_destination · delete_destination

Alert templates: get_alert_templates · set_alert_template

Agent registry: register_agent · list_agents · get_agent · update_agent · delete_agent

Status pages: list_status_pages · create_status_page · update_status_page · delete_status_page

API keys: create_api_key (optional scope: read / write / admin / ingest) · list_api_keys · regenerate_api_key (new secret, same key; does not cascade) · revoke_api_key (cascades to every key it created)

Terraform: export_terraform

This binary carries the same tool set as the hosted server at mcp.lastping.dev. It is a thin REST client throughout: every tool is a direct HTTP call to the management API, so it stays free to run yourself with no lag behind the hosted surface beyond a new release.

The one that matters most is get_ping_instructions: an agent calls create_monitor, then asks for its own ping commands, and wires them into its own work — in one conversation, without a human opening a dashboard.

Ping API

Every monitor gets a URL. There is nothing to install and no library to keep current; anything that can make an HTTP request can report.

What happened	Request
finished successfully	`POST <ping-url>`
started a run	`POST <ping-url>/start`
failed	`POST <ping-url>/fail` with the error as the body
exited with a code	`POST <ping-url>/<exit-code>`
waiting on a human	`POST <ping-url>/blocked`
progress worth recording	`POST <ping-url>/note`

Add ?rid=<id> to pair a run's start with its result, so LastPing can group a run's pings and time it.

# The classic one-liner, at the end of a cron job:
curl -fsS -m 10 --retry 3 https://ping.lastping.dev/<monitor-id>

Traces (ships with the next server release)

POST https://ping.lastping.dev/v1/traces accepts an OTLP/HTTP export (application/x-protobuf or application/json, gzip accepted) with a Bearer key (an ingest key bound to the monitor, from create_ingest_key; a write or admin key also works), or POST <ping-url>/v1/traces for exporters that cannot set headers. Spans need resource attributes lastping.monitor_id and lastping.run_id to be accepted; a payload is capped at 1 MiB decompressed, 500 spans per request and 2,000 spans per run. lastping run sets all of this up for you — see Traces above.

Monitoring as code

resource "lastping_monitor" "nightly_etl" {
  name          = "nightly-etl"
  slug          = "nightly-etl"
  schedule_kind = "cron"
  cron_expr     = "0 3 * * *"
  tz            = "Europe/Berlin"
  grace_s       = 900
}

The provider is on the Terraform Registry as lastping-dev/lastping, with source at lastping-dev/terraform-provider-lastping.

License

MIT. See LICENSE.

LastPing MCP

Документация

Install

lastping run — reporting you can't forget

Traces (ships with the next server release)

MCP server — let an agent set up its own monitoring

Ping API

Traces (ships with the next server release)

Monitoring as code

Links

License

`lastping run` — reporting you can't forget