Skip to main content
Glama

datasieve-mcp

MCP server for DataSieve — standing infrastructure for AI agents, paid over x402. On-chain conditions on Base (wallet activity, whale transfers, price crossings, new pools), url-change web monitoring (CSS selector / regex scoped), dead-man switches for your own agents (your agent pings a URL; silence fires the alert), cron schedules (fire a webhook every N seconds, at a time, or on a cron expression), and multi-agent coordination (barriers + locks). Every event carries a signed Ed25519 observation receipt. No account, no API key: one gasless USDC payment per watch, signed locally — your key never leaves your machine. Failed calls are never billed.

Setup — no key required to start

claude mcp add datasieve -- npx -y datasieve-mcp

Most tools work with no wallet at all — check our scanner's health, read watch status, drain events, ping dead-man switches, pause/resume, verify receipts, cancel watches. Only the buying tools spend money.

npx -y datasieve-mcp --install-skill

Installs a skill into ~/.claude/skills/datasieve/ that teaches Claude when to reach for standing infrastructure proactively: arm a dead-man switch before long autonomous work, schedule wake-ups, watch pages and wallets, coordinate sub-agents.

When you want to buy

x402 means paying is a signature, so purchases need a key. Two rules:

  1. Use a dedicated burner wallet holding a few dollars of USDC on Base — never your main wallet. Payments are gasless (EIP-3009), so it needs no ETH.

  2. Keep the key out of config files — point at a file instead:

echo '0xYOUR_BURNER_KEY' > ~/.datasieve/key.txt && chmod 600 ~/.datasieve/key.txt
claude mcp add datasieve -e DATASIEVE_PRIVATE_KEY_FILE=$HOME/.datasieve/key.txt -- npx -y datasieve-mcp

(DATASIEVE_PRIVATE_KEY=0x... inline also works if you prefer.)

The key never leaves the process: each payment is a signature authorizing exactly the route price — one at a time, nothing standing. Run payment_status to see the wallet, its balance, and prices before spending anything.

Rehearse for free

claude mcp add datasieve-sandbox -e DATASIEVE_API_URL=https://staging.datasieve.xyz \
  -e DATASIEVE_PRIVATE_KEY_FILE=$HOME/.datasieve/key.txt -- npx -y datasieve-mcp

Same API, Base Sepolia, faucet USDC. Sandbox watches last 1 hour.

Related MCP server: @hpp-io/x402-mcp-bridge

Tools

tool

cost

what

create_watch

$0.10 (standard/7d) · $0.25 (composite/7d, ≤5 conditions) · $0.35 (standard30/30d) · $0.85 (composite30/30d)

buy a watch; returns id, secret, ping URLs

renew_watch

tier price

one more period, trigger budget reset

coord_workspace

$0.25 / 7d

multi-agent barriers + locks; every op after is free

cron_next

$0.001

parse a cron expression: next N run times, any timezone

x402_ping

$0.001

end-to-end payment stack self-test

watch_status

free

heartbeat, triggers, expiry, up/down per dead-man switch

drain_events

free

collect everything caught while you weren't running

pause_watch / resume_watch

free

stop firing without deleting; backlog-free resume

ping_heartbeat

free

keep your dead-man switch alive

verify_receipt

free

check the Ed25519 receipt on any delivered event

coord_barrier / coord_arrive / coord_lock / coord_release / coord_events

free

workspace ops

delete_watch

free

cancel

service_health

free

verify our scanner is alive before you pay

payment_status

free

wallet address, USDC balance, prices — "can I buy?"

Watch ids and secrets are remembered in ~/.datasieve/watches.json (mode 600), so follow-up calls only need the watch id — the persistent memory your ephemeral agent lacks.

Available Tools

19 tools
coord_arriveArrive at a barrier (free, no auth — the URL is the capability)A

Report that this part is done. Pass the arriveUrl from coord_barrier plus a part name and optional ctx (whatever you want handed to whoever consumes the released barrier). Idempotent per part: a retrying agent cannot inflate the count.

Inputs: part (required), arriveUrl (required), ctx.

coord_barrierCreate a barrier / fan-in (free — workspace op)A

BARRIER: "wake me when all N of my sub-agents finish." Returns an arriveUrl — hand it to each sub-agent (it is a capability: they can arrive but cannot read or delete your workspace). When the N-th part arrives we fire ONE webhook carrying every part's collected context. If the TTL passes first you get a "partial" event naming exactly who never showed. Timeout is a feature, not an error.

Inputs: parts (required), workspaceId (required), label, secret, +2 more.

coord_eventsDrain coordination events (free) — the mailbox for agents with no serverB

Collect barrier releases/partials that happened while nothing of yours was running.

Inputs: workspaceId (required), since, secret.

coord_lockAcquire a lock / lease (free — workspace op)A

LOCK: "only one agent does this irreversible thing." Returns a leaseId, or 409 with who holds it. The lease AUTO-EXPIRES after ttlSeconds, so a crashed agent can never deadlock your fleet. Call coord_release when done (or just let it expire).

Inputs: key (required), workspaceId (required), holder, secret, +1 more.

coord_releaseRelease a lock (free — workspace op)D

Inputs: key (required), leaseId (required), workspaceId (required), secret.

coord_workspaceBuy a coordination workspace (paid via x402)A

Buy a coordination workspace ($0.25 / 7 days, 1000 ops) — the only x402 seller of multi-agent primitives. After this ONE payment, every barrier and lock operation is FREE and FAST (no per-op settlement). Use it to fan-in sub-agents (coord_barrier) and to stop two agents doing the same irreversible thing (coord_lock).

Inputs: webhookUrl.

create_watchCreate a watch (paid via x402)A

Buy a standing watch from DataSieve (https://api.datasieve.xyz). standard = $0.10/7d, 1 condition, 50 triggers; composite = $0.25/7d, up to 5 any-match conditions (e.g. 5 heartbeat dead-man switches), 200 triggers; standard30 = $0.35/30d and composite30 = $0.85/30d — one payment covers the month, for agents that will not be awake to renew weekly. Payment is a gasless USDC transfer signed locally. Events go to your webhookUrl (HMAC-signed) if set, and always to a server-side mailbox drained with drain_events; every event carries a signed Ed25519 observation receipt you can replay to third parties. Failed calls are never billed. The watch id + secret are remembered locally so later calls need only the id.

Inputs: conditions (required), tier, webhookUrl.

cron_nextParse a cron expression (paid via x402, $0.001)A

Next N run times for a cron expression in any IANA timezone, plus seconds-until-next. If the goal is for something to HAPPEN at those times, create_watch with a schedule condition instead — DataSieve fires the cron as signed webhooks for the life of the watch.

Inputs: cron (required), tz, count.

delete_watchCancel a watch (free, no refund)D

Inputs: watchId (required), secret.

drain_eventsDrain held events (free)C

Collect everything the watch caught — works with zero infrastructure (the mailbox pattern). Pass since = last seen event id for incremental drains.

Inputs: watchId (required), since, secret.

pause_watchPause a watch (free)A

Stop a watch firing (chain, web, heartbeat, schedule) without deleting it. The expiry clock keeps running. Use before a planned agent shutdown so your dead-man switch does not cry wolf.

Inputs: watchId (required), secret.

payment_statusCan I buy? (free) — wallet, balance, pricesA

Check whether a buying key is configured and whether it can afford a watch. Reports the wallet address, its USDC balance on the configured network, and what each tier costs. Call this before create_watch; it never spends anything.

Inputs: no named parameters.

ping_heartbeatHeartbeat check-in (free)A

Keep a dead-man switch alive. Pass the pingUrl from create_watch (or a watchId whose ping URLs are in local state). Call this at least once per intervalSeconds.

Inputs: pingUrl, watchId.

renew_watchRenew a watch (paid via x402)A

Extends the watch one more period (7 days, or 30 on the *30 tiers) at its tier price and resets the trigger budget. Renewing a paused watch also reactivates it.

Inputs: watchId (required), tier.

resume_watchResume a paused watch (free)A

Backlog-free resume: heartbeat deadlines restart from now and schedules jump to their next FUTURE slot — nothing that elapsed during the pause replays.

Inputs: watchId (required), secret.

service_healthDataSieve service heartbeat (free, public)A

Scanner cursor lag vs chain head — verify the watcher is alive before you buy.

Inputs: no named parameters.

verify_receiptVerify a signed observation receipt (free)A

Every DataSieve event carries receipt {alg, keyId, statement, sig} — an Ed25519 attestation that DataSieve observed the event. This checks it against the service key, so an agent can validate evidence handed to it by ANOTHER agent without trusting that agent.

Inputs: receipt (required).

watch_statusWatch status (free)C

State, heartbeat lastCheckedAt, triggersUsed/cap, expiry, delivery stats, and for dead-man switches: up/down, lastPingAt, nextDeadline, pingUrl. Secret is looked up from local state if omitted.

Inputs: watchId (required), secret.

x402_pingTest the payment stack end-to-end (paid via x402, $0.001)A

The cheapest possible real x402 purchase: proves this wallet + client can pay a production seller before you point it at anything expensive. Returns payer + network; failed calls are never billed.

Inputs: no named parameters.

Tool Schema Changelog

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

Loading tool schema changelog…

    TDQS

    B3/5.0

    Scored across 19 tools

    Disambiguation4/5

    Most tools have clearly distinct purposes: watch lifecycle (create/pause/resume/renew/delete/status), coordination (barrier/arrive/lock/release/workspace), and payment/health (x402_ping/payment_status/service_health). Minor overlap exists between drain_events and coord_events (both collect events), and coord_release lacks a description, causing slight ambiguity with coord_lock's release.

    Naming Consistency3/5

    Naming is a mix of verb_noun (create_watch, pause_watch, delete_watch), noun_verb (watch_status, service_health), and prefixed coord_* tools. Each subdomain is internally consistent, but there is no single pattern across the entire set, making it predictable only within groups.

    Tool Count4/5

    19 tools is above the ideal 3-15 range but justified by the broad scope: watch management, coordination primitives, payment validation, and event verification. Each tool appears necessary, though a few (e.g., two with no descriptions) could be consolidated.

    Completeness4/5

    The watch lifecycle is fully covered (create, read via status, pause, resume, renew, delete), and event draining and receipt verification are included. Coordination barrier and lock patterns are complete. Minor gaps exist, such as no list_watches or ability to modify watch conditions, but core workflows are solid.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers