Pocket Agents MCP Server

pocket-agents

Seven verification-style services for the Pocket Network Agentic Portal, built and operated by one supplier on a Mac mini. Stdlib-only Python (3.9+), no keys, every response a JSON object, deterministic per snapshot, facts separated from judgments.

All six pass the portal audit 9/9 on Beta TestNet (2026-09-20) with settled claims served through the new pocket-relay-miner.

Use from agents and CI

• MCP: every service as Model Context Protocol tools. Remote: https://mcp.pokt-agent.com/mcp (no sign-in). Local stdio server: mcp/.
• GitHub Action: check the Pine files in a repository on every push — Pine Check on the GitHub Marketplace: uses: jsymen1290/pine-check-action@v1.
• Also live since 2026-10-05 (REST + Pocket Beta): Backtest Reconcile, KR Figure Check, DART Correction Impact, KR Rates & FX, KR Law, KR Apartment Trades.

Two calls, shown

Both of these are live right now, no key and no account.

Does this Pine script make backtests look better than live trading?

curl -s -X POST https://pine.pokt-agent.com/v1/integrity -H 'content-type: application/json'   -d '{"source":"//@version=6
indicator(\"x\", overlay=true)
d = request.security(syminfo.tickerid, \"D\", close, lookahead=barmerge.lookahead_on)
plot(d)
"}'

{"verdict": "FAIL", "counts": {"error": 1, "warning": 1, "info": 0},
 "findings": [
   {"code": "P007", "severity": "warning", "line": 3, "message": "request.security with lookahead_on repaints on historical bars"},
   {"code": "P201", "severity": "error",   "line": 3, "message": "request.security(..., lookahead_on) without a [1]-style offset reads the future on historical bars"}]}

Add close[1] to that same call and the error disappears: the accepted confirmed-bar idiom is recognised, not punished.

What would 0.5 BTC actually cost on a Korean exchange right now?

curl -s 'https://kr.pokt-agent.com/v1/executable/BTC?qty=0.5&side=buy&fee_pct=0.05'

{"venue": "upbit", "average_price_krw": 116381997.41, "coverage": 1.0,
 "slippage_pct": 0.0009, "orderbook_lag_seconds": 1.865}

coverage is the fraction of the quantity the visible book can fill, slippage_pct is against best price, and orderbook_lag_seconds is how stale the book was when we read it. A venue that is down returns a status code, never a guess.

The same walk answers a plain question in English or Korean:

curl -s -X POST https://kr.pokt-agent.com/v1/query -H 'content-type: application/json'   -d '{"question":"sell 2 BTC on Upbit and Bithumb, what fills?"}'

On 2026-09-28 that returned coverage 0.77 on Upbit and 0.50 on Bithumb: the visible books could not fill 2 BTC at any price.

Median response time measured from the host on 2026-09-22: 248-333 ms across the six services. Services that call an upstream API (dart-kr-events-v1, kr-export-pulse-v1) replay their own card probes against themselves just inside the cache TTL (warmup.py), so the first call after a restart is 0.04 s instead of 3.8 s.

Layout

pocket_agents/
  client.py        GET-only allow-listed HTTP client for public Pocket REST/RPC (budgeted)
  indexer.py       data.pocket.network GraphQL: every settlement for an operator and range, complete coverage
  chain.py         LCD/RPC queries: params, supplier, service, balance, block_search/results
  settlement.py    EventClaimSettled extraction + aggregation
  collect.py       always-on collector (STATE/ALERTS/settlements ndjson) that the agents read
  surveillance.py  15 official sources: fetch, diff, receipts, verify
  economics.py     supplier/network economics from public analytics (observed values only)
  ops.py           operator watch (facts vs judgments)
  agent.py         pokt-settlement-agent-v1 HTTP service (:8793) + service card + OpenAPI
  watch.py         pokt-network-watch-v1 (:8794)
  krmarket.py      kr-market-data-v1 (:8795)
  pinelint.py      pine-script-lint-v1 (:8796)
  benchmark.py     agent-service-benchmark-v1 (:8799) portal catalogue + acceptance audit comparison
  krexport.py      kr-export-pulse-v1 (:8797)  needs DATA_GO_KR_SERVICE_KEY
  dartevents.py    dart-kr-events-v1 (:8798)   needs OPENDART_API_KEY
  cli.py           entry points: *-serve, *-card, collect, surveil, ask
tests/test_services.py   offline tests with fixtures (no network)
deploy/            watch list, surveillance sources, RelayMiner/launchd templates (no secrets)
docs/              design notes

Run locally

python -m unittest tests.test_services
python -m pocket_agents pinelint-serve --port 8796     # then POST http://127.0.0.1:8796/v1/lint {"source": "..."}
python -m pocket_agents krmarket-serve --port 8795
python -m pocket_agents watch-serve --port 8794        # needs data/ written by `collect` and `surveil`
python -m pocket_agents agent-serve --port 8793 --no-llm

Optional LLM narration in the settlement agent and surveillance review is disabled unless a runtime with config/router modules is present; every answer has a deterministic template fallback and a number guard.

Serving on Pocket

Each service ships its on-chain card (*-card) and OpenAPI. Deployment follows the portal's design rules: JSON-object responses, inputs in the body, no caller auth, no error-looking words in success bodies, identity/readiness/functional probes, determinism. See deploy/ for the relayer/miner config shapes and docs/ for the registration record (service, supplier, application stake, first settled claim).

License

MIT