io.github.Shxnque/quesen MCP Server

io.github.Shxnque/quesen is an MCP server that provides a deterministic decision layer for autonomous agents. It defines reproducible PROCEED, REVIEW, and SKIP verdicts for A2A (Agent-to-Agent) risk evaluation, supporting agent risk management with an agent-guardrails and agent-safety focus.

🛠️ Key Features

Deterministic decision layer for autonomous agents
Reproducible PROCEED, REVIEW, SKIP verdicts
A2A risk evaluation and agent risk management
MCP compatibility (per server README)

🚀 Use Cases

Agent-to-agent (A2A) risk evaluation
Guardrails for autonomous-agent workflows
Risk scoring / risk management in LLM agents

⚡ Developer Benefits

Deterministic, reproducible decision outcomes
Clear verdict categories: PROCEED, REVIEW, SKIP
Built for agent infrastructure and agent safety use

⚠️ Limitations

Tool set details are not provided beyond a stated tool count (toolCount: 5)
No further behavior or API specifics included in the provided excerpt

Quesen — Developer Portal

Quesen is the deterministic decision-and-receipt core for agent actions — a typed security context in, a PASS / REVIEW / BLOCK / SKIP verdict out, with machine reason codes and a receipt you can re-run byte-for-byte and prove. No model inference is in the scoring path, so the same input always yields the same verdict. It is built to sit behind injection detection, on top of agent identity, and to bill per decision (ASP/402).

Unlike log-based governance layers whose audit trail is their word, kept by them, a Quesen receipt is independently verifiable by the caller — recomputable, and (engine signing enabled) Ed25519-signed. See docs/architecture-gap-closers.md and client-side enforcement + receipt verification in quesen-sdk ≥ 0.5.0. Production engine signing is live — verify a real signed receipt yourself in docs/verify-live-receipt.md (no trust in Senueren required).

This repository is the public developer portal. It contains only documentation, integration guides, examples, registry manifests, and reference links. No engine source code lives here. Quesen's engine implementation is sovereign, non-public infrastructure.

Live production

Surface	URL
REST API	`https://web-production-3df26.up.railway.app`
MCP (Streamable HTTP)	`https://web-production-3df26.up.railway.app/mcp`
OpenAPI 3.1	`https://web-production-3df26.up.railway.app/openapi.json`
Swagger UI	`https://web-production-3df26.up.railway.app/docs`
Health	`https://web-production-3df26.up.railway.app/health`
Version	`https://web-production-3df26.up.railway.app/version`

Quick start (30 seconds)

Fastest path — no install, no signup, no card. Self-serve a free sandbox key and run a real deterministic decision against production. Full guide: docs/QUICKSTART.md · try it in the browser at senueren.co.za/try.

bash

# 1 · get a free sandbox key
curl -X POST https://web-production-3df26.up.railway.app/sandbox/keys

# 2 · evaluate an action (use the api_key from step 1)
curl -X POST https://web-production-3df26.up.railway.app/validate \
  -H "X-API-Key: sk_sandbox_..." \
  -H "Content-Type: application/json" \
  -d '{"domain_age_days": 1, "engagement_ratio": 0.95, "scam_keyword_count": 4}'
# -> {"decision":"SKIP","risk_score":1.0,"conflict_triggers":[...],"input_snapshot_hash":"..."}

SDKs

Published. The SDKs are live on PyPI and npm (quesen-sdk 0.5.0 / npm 0.5.0; quesen-langchain, quesen-crewai, quesen-autogen 0.3.0). The base_url + X-API-Key (including the sandbox key above) are identical across all SDKs.

Python

bash

pip install quesen-sdk   # PyPI: https://pypi.org/project/quesen-sdk/

python

from quesen_sdk import QuesenClient

q = QuesenClient(base_url="https://web-production-3df26.up.railway.app",
                 api_key="YOUR_KEY")   # a sandbox key from /sandbox/keys works here

verdict = q.validate(domain_age_days=1, engagement_ratio=0.95, scam_keyword_count=4)
if verdict.decision == "SKIP":
    return  # respect the deterministic answer

JavaScript / TypeScript

bash

npm i quesen-sdk   # npm: https://www.npmjs.com/package/quesen-sdk

import { QuesenClient } from "quesen-sdk";

const q = new QuesenClient({
  baseUrl: "https://web-production-3df26.up.railway.app",
  apiKey: process.env.QUESEN_API_KEY,
});

const verdict = await q.validate({
  domain_age_days: 1,
  engagement_ratio: 0.95,
  scam_keyword_count: 4,
});

Framework wrappers

Framework	Package	Repository
LangChain / LangGraph	`quesen-langchain`	Shxnque/quesen-langchain
CrewAI	`quesen-crewai`	Shxnque/quesen-crewai
AutoGen v0.4+	`quesen-autogen`	Shxnque/quesen-autogen
Python (core)	`quesen-sdk`	Shxnque/quesen-sdk-py
JavaScript / TypeScript	`quesen-sdk` (npm)	Shxnque/quesen-sdk-js

MCP (Claude Desktop, Cursor, Windsurf, etc.)

Quesen exposes five MCP tools over the production endpoint. See docs/mcp.md for the client-config snippet.

Why Quesen?

Autonomous agents make more decisions per second than any human oversight can audit. When those decisions involve capital — launching a token, opening a position, executing a trade, greenlighting a smart-contract deployment — the marginal cost of a bad decision is fatal.

Quesen answers exactly one question:

Should the calling agent proceed with this action?

Inputs are typed. Outputs are one of PROCEED, REVIEW, SKIP, always with a risk_score in [0.0, 1.0], a confidence in [0.0, 1.0], and the exact conflict rules that fired. Same inputs → same output. Every time. Every response embeds engine_version, weights, and thresholds. Fully reproducible. Fully auditable.

What Quesen is not

Not an LLM wrapper. No prompts. No probabilities.
Not a chatbot. It is A2A infrastructure.
Not a KYC/identity system. It scores risk, not identity.
Not chain-locked / framework-locked / LLM-locked. Ecosystem-neutral by design.

Documentation

Quickstart — first decision in under 10 minutes (free sandbox key).
Architecture overview
Integration guide
API reference
MCP setup
Pricing tiers
FAQ
Registry status
Evidence & case studies — verified upstream merges + live, byte-for-byte replayable receipts. Engineering/standards evidence only; no customer traction is claimed.

Independent verification

Published receipts are independently reproducible from this repo alone — no hosted service or private engine required:

bash

python3 verify/verify_receipts.py          # offline, stdlib-only
python3 verify/verify_receipts.py --live   # also cross-check the live engine

All six UCP #724 vectors show a byte-for-byte three-way match between the published fixture, the public reference, and the live engine (verify/README.md, verify/three_way_match.json). That doc also states honestly where independent verification stops today (the production ruleset commit_sha is not publicly resolvable; receipts are not yet cryptographically issuer-signed).

The egress/authority decision subset — the part security integrators gate on — is now independently verdict-replayable offline too, with zero network:

bash

python3 evaluation/conformance/verify_conformance.py   # offline; recomputes decision+reasons+hash

Six cases (OWASP-agentic + LoopX prepared-Effect PASS/REVIEW/BLOCK) recompute byte-for-byte from the public reference evaluator, plus a prod-1→prod-2 integrity-flip check — no signup, key, or hosted call (evaluation/conformance/README.md).

Tutorials

Moltbook post-guard — deterministic pre-post safety gate for autonomous social agents.
OpenClaw MCP plugin — wiring Quesen as an MCP-native guardrail into OpenClaw-style agents.

Live status

Production: https://web-production-3df26.up.railway.app
Health check: GET /health returns {"status":"ok","engine_version":"1.10.0"}
Version snapshot: GET /version returns full engine + billing + on-chain flags (ASP/1.0)
Uptime and version widget on senueren.co.za/quesen

Registry presence

Quesen is discoverable via Model Context Protocol registries and the standard agent-directory ecosystem. See docs/registries.md for the current state of each submission. Manifests:

smithery.yaml — Smithery.ai (canonical)
mcp.json — MCP.so / generic MCP client (canonical)
.well-known/ai-plugin.json — OpenAI plugin manifest / .well-known/ai-plugin.json autodiscovery
llms.txt — machine-readable summary for LLM crawlers

Contributing

This is a documentation-only repository. Engine PRs cannot be accepted here. If you have integration-specific feedback, please open an issue or read CONTRIBUTING.md.

SDK contributions belong in the corresponding public SDK repository:

Security issues: please read SECURITY.md before filing publicly.

Community

Follow build updates, agent-governance notes, and Quesen releases on X: @SenuerenGroup · web: senueren.co.za · senueren.co.za/quesen.

If a Quesen integration or a merged contribution helped your project, a follow or a note at @SenuerenGroup is always appreciated.

License

MIT. See LICENSE.

domain_age_days	integer \| null	optional	Age of the target domain in days.
engagement_ratio	number \| null	optional	Engagement ratio in [0, 1].
scam_keyword_count	integer \| null	optional	Count of known scam keywords detected.
client_request_id	string \| null	optional	Optional caller-side trace id (echoed in response).

domain_age_days	integer \| null	optional	Age of the target domain in days. Same feature as quesen.validate; combined with the (possibly overridden) weights to produce a counterfactual risk score.
engagement_ratio	number \| null	optional	Engagement ratio in [0, 1]. Same feature as quesen.validate; used with the (possibly overridden) weights during simulation.
scam_keyword_count	integer \| null	optional	Count of known scam keywords detected. Same feature as quesen.validate; used with the (possibly overridden) weights.
weights_override	object \| null	optional	Optional per-feature weight override used ONLY for this simulation. When null (or omitted) the engine uses its production weights. Does NOT mutate global state.
thresholds_override	object \| null	optional	Optional decision-threshold override used ONLY for this simulation. When null the engine uses its production thresholds. Does NOT mutate global state.
client_request_id	string \| null	optional	Optional caller-side trace id. Echoed verbatim in the response for cross-system correlation.

request_id	string	required	The `request_id` returned by a prior quesen.validate call. Ties the reported outcome back to the exact decision that produced it.
outcome	string	required	Real-world outcome once known. RUG = rug-pull / total loss; LOSS = negative PnL; OK = neutral / no material result; WIN = positive PnL; UNKNOWN = outcome not yet observable.
notes	string \| null	optional	Optional free-text notes about the outcome (post-mortem, context, references). Stored verbatim for audit.
realized_pnl	number \| null	optional	Realized profit/loss in the caller's chosen unit (e.g. USD, SOL, ETH). Sign convention: positive = profit, negative = loss. Fed into the aggregate `pnl_mean` counter.
elapsed_seconds	integer \| null	optional	Seconds elapsed between the original quesen.validate decision and the observation of this outcome. Non-negative.
venue	string \| null	optional	Venue or platform where the trade/signal was executed (e.g. 'raydium', 'jupiter', 'uniswap-v3'). Free-form label for downstream grouping.
signal_hash	string \| null	optional	Optional hash/fingerprint of the original signal (URL, token address, message id) so outcomes can be joined across systems that don't share Quesen request_ids.
client_request_id	string \| null	optional	Optional caller-side trace id. Echoed verbatim in the response for cross-system correlation.

Quesen Deterministic Agent Decision Layer

io.github.Shxnque/quesen MCP Server

🛠️ Key Features

🚀 Use Cases

⚡ Developer Benefits

⚠️ Limitations

Topics

quesen.validate

quesen.simulate

quesen.report

quesen.health

quesen.version

Quesen — Developer Portal

Quick start (30 seconds)

SDKs

Python

JavaScript / TypeScript

Framework wrappers

MCP (Claude Desktop, Cursor, Windsurf, etc.)

Why Quesen?

What Quesen is not

Documentation

Independent verification

Tutorials

Live status

Registry presence

Contributing

Community

License

Install

Remote endpoint

Related servers