Back to Discover

kaval-clients

connector

LufeMC

Verify the facts an agent's action depends on before it acts — ALLOW, REVIEW, or BLOCK.

View on GitHub
0 starsSynced Aug 5, 2026

Install to Claude Code

/plugin marketplace add LufeMC/kaval-clients

README

Kaval clients

Open-source client libraries for Kaval. Before an agent acts, send Kaval the action. Kaval identifies the facts that action depends on, checks them against the sources it watches, and answers ALLOW, REVIEW, or BLOCK with a signed receipt.

Policy engines decide whether an action is permitted under the rules; Kaval verifies whether the facts those rules depend on are still true.

These are thin HTTP clients for the hosted Kaval API (https://api.usekaval.com). Create an API key at usekaval.com.

PackageLanguageInstallSource
@usekaval/kavalNode / TypeScriptnpm i @usekaval/kavalsdks/node
kavalPythonpip install kavalsdks/python
@usekaval/mcpMCP servernpx -y @usekaval/mcppackages/mcp

0.6 is a breaking release. Nine MCP tools collapsed to seven, and the whole verification surface collapsed to one call. Every removed endpoint now answers a structured 410 {"error":"tool_retired","replacement":"/v1/check"}, and the clients translate that into an error that names check by name. See Migrating from 0.5.

One call

import { Kaval } from "@usekaval/kaval";

const kaval = new Kaval({ apiKey: process.env.KAVAL_API_KEY });

const result = await kaval.check({
  action: "Approve this prior-authorization request at the in-network rate",
  context: "payer: Aetna; CPT 12345; plan HMO",
  materiality: "critical",
});

if (result.decision !== "ALLOW") {
  // REVIEW is never permission to act.
  holdForHuman(result.facts.filter((fact) => fact.status !== "holds"));
}
import os

from kaval import KavalClient

kaval = KavalClient(api_key=os.environ["KAVAL_API_KEY"])
result = kaval.check(action="Approve this prior-authorization request at the in-network rate")
if result["decision"] != "ALLOW":
    hold_for_human(result["facts"])

What comes back:

fieldmeaning
decisionALLOW (every material fact holds on fresh evidence) · REVIEW · BLOCK
reason_codeswhy, from a closed eight-code taxonomy
facts[]one row per fact: status (holds/changed/unknown), materiality, and the sources it rests on
receipt{ id, signature, signed_at } — fetch the full signed document with getReceipt(id)
latency_ms{ compile, lookup, live, total }

A check on facts a watched source already covers is a database read and returns in tens of milliseconds. A cold check does live research before it answers — search, fetch, adjudicate — and the server lets that run for up to 100s by default, so give the call room. mode: "fast" (equivalently max_wait_ms: 0) skips research entirely and reports anything it could not settle as unknown, which is REVIEW.

The decision table is published, so the receipt's fact list re-derives the verdict offline, and the Ed25519 public keys are served unauthenticated at GET /v1/proof-verification-keys/:kid — checking a receipt needs no Kaval account and no API key.

The verifier that does it for you ships inside the SDK: @usekaval/kaval/verify is a dependency-free subpath export of @usekaval/kaval, and the same package ships a kaval-receipt-verify CLI. Neither needs a Kaval account, an API key, or Kaval's database; the only request either can make is for the public keyset, and that request is optional. Hand it a receipt and a keyset. It answers three questions by default and one optional verdict question:

  1. Cryptographic validity — does the Ed25519 signature cover the exact canonical unsigned bytes?
  2. Key trust — is that key_id active or benignly retired, rather than revoked or compromised?
  3. Freshnessfresh, recheck_due, expired, not_yet_issued, or unknown.
  4. Verdict derivation — does the receipt's fact list produce its stated verdict and reason codes?

A valid signature proves who sealed those exact bytes. It does not prove the claim is still true, or that the key is still trusted, which is why the three answers never collapse into one boolean.

import { extractReceipt, parseJsonStrict, verifyReceipt } from "@usekaval/kaval/verify";

const receipt = extractReceipt(parseJsonStrict(receiptText));
const result = verifyReceipt(receipt, parseJsonStrict(keysetText), { derive_verdict: true });

result.cryptographic.valid; // the signature covers these exact canonical bytes
result.key.trusted;         // the signing key is not revoked or compromised
result.freshness.status;    // separate fact — a check receipt carries no expiry, so `unknown`
result.decision?.matches;   // the published table reproduced the verdict and reason codes
# Reproducible audit: archive the keyset beside the receipt and stay entirely offline.
npx -p @usekaval/kaval kaval-receipt-verify verify receipt.json --keyset keys.json
npx -p @usekaval/kaval kaval-receipt-verify verify receipt.json --keyset keys.json --derive-verdict

# Or resolve the key over HTTPS from the unauthenticated endpoint.
npx -p @usekaval/kaval kaval-receipt-verify verify receipt.json \
  --key-url https://api.usekaval.com/v1/proof-verification-keys

Exit 0 means the signature is valid and the key is trusted; a stale receipt still exits 0, because freshness is a separate fact — pass --require-fresh to make anything but fresh non-zero. Exit 1 is a completed but unaccepted verification, 2 an input, I/O, or discovery failure. Parse untrusted receipt text with parseJsonStrict, not JSON.parse: duplicate members and lossy numbers are evidence, and JSON.parse throws that evidence away before any verifier can see it.

Keep it warm: watch the sources

A check is a database read when the facts it needs are already backed by a watched source, and a bounded research run when they are not. Registering the name of an authority is usually enough:

await kaval.addSource({
  kind: "entity",
  name: "Aetna",
  intent: "payer policy bulletins",
});

Kaval resolves that to the pages that publish it, polls them adaptively (slower when nothing changes, faster when it does), and re-evaluates the dependent facts when they move. kind: "url" watches one page; kind: "push" is a document your own system sends in with sendEvent(). You do not have to register first — a source a check cites is auto-watched — but registering ahead of time is what makes the first check on a fact fast.

Naming an entity is what enqueues the discovery that works out how to acquire the pages behind it. A kind: "url" source registered directly does not get one, so recompileSource(id) is how you ask for one — and it is also the only way back once a source's acquisition plan breaks. It answers 202 { source_id, job_id, created }; created: false means an open job already covered it.

Close the loop: subscribe to deltas

Watching is only half of it. Subscribe to fact_state.delta and Kaval pushes you what changed and what it flipped, instead of you discovering it on the next check:

const { webhook_verification } = await kaval.subscribeFactStateDeltas({
  callback_url: "https://your-app.example.com/hooks/kaval",
  external_scope_ids: ["plan:HMO"], // optional filter
});
// Store webhook_verification.secret — it is shown exactly once, and it is how you
// authenticate every inbound delivery.

Each delivery names the source, the old and new content hashes, a diff summary, and the facts whose state changed — {fingerprint, text, old_state → new_state, basis} — plus a pointer to the receipt covering the re-evaluation. Manage subscriptions with listWebhooks(), setWebhookEnabled(), deleteWebhook(), and re-drive a dead letter with replayWebhookDelivery().

Push your own documents

For documents Kaval cannot fetch — a contract, an internal policy, a customer upload — push the new version and let the background loop do the rest:

const { changed, facts_pending_review } = await kaval.sendEvent({
  namespace: "contracts",
  document_id: "msa-2026-07",
  content: extractedText,
  scope_keys: ["contract:msa-2026-07"],
});

Kaval diffs it against the previous version, marks the dependent facts stale, re-evaluates them, and emits the delta webhook. Checks that land mid-re-evaluation honestly return REVIEW.

MCP

KAVAL_API_KEY=kv_live_… npx -y @usekaval/mcp

Seven tools over stdio: check (the one that does the work), get_receipt (the full signed document behind a verdict), add_source / list_sources / remove_source, report_outcome, and the deprecated verify pilot alias. See packages/mcp.

MCP clients cancel a tool call well before 100s, so the server narrows check's research budget to fit inside that envelope instead of inheriting the full default. A cold check over MCP therefore comes back REVIEW more often than the same call through an SDK — register the sources it depends on and the warm path removes the difference.

Migrating from 0.5

Everything below folded into check. The old routes answer 410 tool_retired; the clients raise KavalRetiredError (Node) / KavalRetiredError (Python) naming the replacement, and the MCP server returns {"error":"tool_retired"} with a message telling the agent to call check.

MCP tools

0.5 tool0.6
currentness_checkcheck{ action } or { claims: ["…"] }
currentness_verifycheck — branch on decision === "ALLOW" instead of act
currentness_extract_and_checkcheck — pass the paragraph as action/context; Kaval compiles the facts itself
currentness_scan_storecheck{ claims: [...] }, up to 20 per call
currentness_monitoradd_source + a fact_state webhook subscription — deltas are pushed, not swept
proof_auditcheck — the receipt is the proof; get_receipt fetches it in full
proof_gatecheck — the warm path re-checks in ~50ms, so there is nothing to re-apply separately
report_outcomereport_outcome (unchanged; pass receipt.id)
verifyverify, now deprecated → move to check

Node / Python methods

0.50.6
audit(), gate(), gateAction()/gate_action()check()
check(belief) / check(belief=…)check({ action }) / check(action=…)
verifyBelief() / legacy_verify_belief()check({ action, context })
extractAndCheck() / extract_and_check()check({ action: theText })
scanStore() / scan_store()check({ claims: [...] })
monitor()addSource() + subscribeFactStateDeltas()
kaval() / kavalBatch() / kaval_batch()check({ claims: [{subject, predicate, object, scope}] })
verify()verify(), deprecated → check()
new: getReceipt, listSources, recompileSource, sendEvent, webhooks
ProofNotFoundError / KavalProofNotFoundErrorremoved with /v1/gate

Verdict mapping

0.5 belief status0.6
current + act: truedecision: "ALLOW", every fact holds
stale / contradictedfact changedREVIEW or BLOCK by materiality
unsupported / insufficient / conflictingfact unknownREVIEW (BLOCK if critical)

Idempotency

verify() (the deprecated pilot alias) and webhook creation are the only client operations that carry an Idempotency-Key. A check is a read of current state — the server does not replay it, so retrying one is free and no key is spent. verify() keeps the bounded single retry for ambiguous transport outcomes; createWebhook() generates a key when you do not supply one, because the API requires one.

API origin env vars

Two names exist on purpose — they are not interchangeable:

ConsumerVariableReads env?
Python SDK / MCPKAVAL_BASE_URLyes
Node @usekaval/kavalpass baseUrl in constructor
Marketing site proxy (apps/web)KAVAL_API_URLyes (server only)

Use the same origin value in both vars when self-hosting (e.g. http://localhost:8787, the port the server image listens on). The self-host guide ships with the server distribution rather than here — Kaval's core repo is private, so there is no public link to give you.

Honest boundaries

Demo results carry no organizational authority; a production ALLOW requires a customer-bound action policy and applicable empirical calibration; REVIEW is never permission to act.

Development

pnpm install
pnpm check        # build + lint + typecheck + test (the JS packages)
pnpm check:docs   # this README and the CHANGELOG against the shipped surface

# Python SDK
cd sdks/python && pip install -e ".[dev]" && pytest

Everything above is hermetic: it fakes the API, which is why a client that aborts its own headline call, or an endpoint that 404s, can pass it. Two suites talk to a real server, and both skip themselves when their credentials are absent:

export KAVAL_API_KEY=kv_live_…
export KAVAL_BASE_URL=https://api.usekaval.com   # or http://localhost:8787

pnpm --filter @usekaval/mcp exec vitest run test/live-tools.test.ts
cd sdks/python && pytest tests/test_live.py

These need a running Kaval server and an issued API key, so they are opt-in and self-skip without one. They are not a release gate here: this repository cannot reach the server's source, and there is no staging deployment — the only deployment is production, and pointing a publish gate at it would write test sources, receipts and outcome reports into the live product on every tag.

The real client-vs-server contract test lives in the Kaval server repository's CI, where a real Postgres and the server both exist. It boots the server, issues a scoped key, and drives these same clients against it on every push. Publishing from this repository gates on the hermetic suites, which is what this repository can honestly verify on its own.

License

Apache-2.0.

Rendered live from LufeMC/kaval-clients's GitHub README — not stored, always reflects the source repo.

1 Install Method

NameDescriptionCategorySource
npm packageInstall via npm (stdio transport)mcp-server@usekaval/mcp

0 Comments

Login required
Log in to post a comment or update on this repo.

No comments yet — be the first to share an update.