Cyan Eclipse Agent Wallet MCP
A sandbox-only agent wallet MCP that lets a human grant scoped spending authority to an AI agent, with policy enforcement, human approval, an audit trail, and an instant kill switch (simulated test USDC — no real funds).
Create wallets owned by a named human; wallets can't spend until a policy is granted, and can optionally be bound to a desk identity (secp256k1 key) so every spend requires a signed authorization with a strictly increasing nonce.
Inspect a wallet: balance, frozen state, today's spend, and active policy.
Grant a spending policy (owner): per-transaction cap, daily cap, approval threshold, recipient allowlist, expiry, and purpose — returning a plain-language liability receipt.
Fund wallets with simulated test USDC via a sandbox faucet (not real money).
Dry-run payments (
check_payment) to get allow / approve / deny with the reason before spending.Execute payments (
pay) inside the granted scope; out-of-scope payments are rejected or queued for human approval if they only exceed the approval threshold.Resolve queued payments: owner approves (executes immediately) or denies pending requests.
Read the append-only audit log of policy grants, payments, rejections, approvals, and revocations (optionally filtered by wallet, with a limit).
Kill switch (
revoke) that freezes the wallet instantly and timestamps the revocation in the audit log.Drive it from any MCP client over stdio (stdlib-only Python), with state persisted in a local JSON ledger.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Cyan Eclipse Agent Wallet MCP check if a $30 payment to 0xMerchantA would be approved"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Eclipse Agent Wallet MCP
Powered by Cyan Eclipse Technologies — https://cetech.ai Listed on Eclipse Bazaar (eclipsebazaar.com).
Why this exists
On November 24, 2026, Wells Fargo's Online Access Agreement Section 15 takes effect: anything your SI agent does with your bank account is legally you — "at your own risk," including transactions you did not want or intend, with no controls attached and no duty on the bank to watch.
This MCP server is the opposite design: an agent wallet where responsibility follows scope.
The human owner grants a written policy: per-transaction cap, daily cap, approval threshold, recipient allowlist, expiry, purpose.
The agent can spend only inside that scope. Outside it, payments are rejected — or routed to a human approval queue when they exceed only the approval threshold.
Every grant, payment, rejection, approval and revocation is written to an append-only audit log — the evidence of what was actually authorized.
A kill switch freezes the wallet instantly, with the revocation timestamped in the log. Revoking access to your bank means "notify us and wait"; here it is immediate and provable.
Related MCP server: wwall
Status
v0.2.0 reference implementation. Sandbox only — balances are simulated test USDC in a local JSON ledger. No real funds move and no network calls are made. The policy engine is the product; production payment rails (x402, partner processors) plug in behind it without changing the tool surface.
Desk extension (v0.2.0, 2026-10-27): a wallet may be bound at creation to a desk identity — a per-desk secp256k1 key held in envelope custody outside the server (desk_keys.py / desk_envelope.py, envelope custody). A desk-bound wallet takes the key's real EVM address, and every spend must carry that desk key's signature (EIP-191 over the canonical spend message; strictly increasing nonce per wallet), verified server-side in pure Python (desk_verify.py). Unbound wallets behave exactly as v0.1.0. Operator console: desk_cli.py (issue / policy / fund / signed pay / revoke). The v0.1.0 policy engine was CERTIFIED(sandbox) by the Eclipse Bazaar Test Bench on 2026-10-24; the desk extension is benched separately (see the team wallet registry record).
Tools (10)
Tool | What it does |
| Create a wallet owned by a named human. Cannot spend until a policy is granted. |
| Balance, frozen state, today's spend, active policy. |
| Owner grants the spending scope; returns a plain-language liability receipt. |
| Sandbox faucet (simulated test USDC) for demos/tests. |
| Dry-run: returns allow / approve / deny with the reason. Agents call this before |
| Execute inside scope; reject outside it; queue for approval above the threshold. |
| Owner resolves a queued request. |
| The full append-only trail. |
| Kill switch — freeze instantly. |
Run it
python3 server.py # MCP over stdio, stdlib only, no dependenciesState persists in ledger.json beside the server (override with AGENT_WALLET_LEDGER).
Use it with your Muse (custom connector / local MCP)
Point your agent's MCP configuration at the command python3 /path/to/server.py — any MCP client (Muse, Muse, Cursor) can drive it. A worked example session:
create_wallet {owner_id: "you@example.com"}→ wallet created, cannot spend.set_policy {per_tx_cap_usdc: "50.00", daily_cap_usdc: "120.00", approval_above_usdc: "20.00", allowlist: ["0xMerchantA"], purpose: "API credits"}→ liability receipt issued.pay 10.00 to 0xMerchantA→ executed.pay 60.00→ rejected (over cap).pay 30.00→ human approval required.revoke→ frozen, timestamped. Agent's authority ends the second the owner says so.
Roadmap
v0.2: hosted HTTP transport + per-user API keys; policy templates (shopping, SaaS, research budgets).
v0.3: real rails behind the same policy engine — x402/USDC settlement, partner-processor fiat — only under the owner's granted scope.
v1.0: Eclipse Bazaar listing with live test-bench scorecard (policy-engine conformance suite runs against every release).
License / attribution
A Cyan Eclipse Technologies product. The "Powered by Cyan Eclipse Technologies" attribution travels with it.
Available Tools
10 toolsapprove_paymentB
Owner approves a pending payment request; it executes immediately.
Inputs: request_id (required).
audit_logC
The append-only audit trail: policy grants, payments, rejections, approvals, revocations. The evidence of what was authorized.
Inputs: limit, wallet_id.
check_paymentA
Dry-run a payment against the policy. Returns allow / approve / deny with the reason. Agents should call this before pay.
Inputs: to (required), wallet_id (required), amount_usdc (required).
create_walletA
Create a sandbox agent wallet owned by a named human. A wallet cannot spend until its owner grants a policy. Optional desk binding (desk_id + desk_pubkey): the wallet takes the desk key's real EVM address and every spend then requires that desk key's signature.
Inputs: owner_id (required), label, desk_id, desk_pubkey.
deny_paymentC
Owner denies a pending payment request.
Inputs: request_id (required).
fund_walletB
SANDBOX faucet: add simulated test USDC so demos and tests can run. Not real money.
Inputs: wallet_id (required), amount_usdc (required).
get_walletB
Wallet summary: balance, frozen state, spend today, active policy.
Inputs: wallet_id (required).
payB
Execute a payment if it is inside the granted scope. Outside scope: rejected, or routed to the human approval queue if it only exceeds the approval threshold. Desk-bound wallets must pass authorization {nonce, signature} from the bound desk key.
Inputs: to (required), wallet_id (required), amount_usdc (required), memo, +1 more.
revokeB
KILL SWITCH: freeze the wallet instantly. All spending stops; the revocation is timestamped in the audit log.
Inputs: wallet_id (required), reason.
set_policyB
The owner grants the spending scope: per-transaction cap, daily cap, approval threshold, recipient allowlist, expiry, purpose. Returns a plain-language liability receipt. Responsibility follows this scope.
Inputs: wallet_id (required), daily_cap_usdc (required), per_tx_cap_usdc (required), purpose, +3 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
Each tool maps to a distinct action on a single resource: create/get wallet, set policy, fund, dry-run check, execute pay, approve/deny pending requests, audit trail, and revoke. The check_payment vs pay split is explicitly clarified as dry-run vs execute, and approve/deny are clearly owner-side actions on pending requests. No two tools appear to do the same thing.
Most tools follow a consistent verb_noun pattern (create_wallet, get_wallet, set_policy, fund_wallet, check_payment, approve_payment, deny_payment). Minor deviations exist with bare verbs (pay, revoke) and a noun-like name (audit_log), but the convention is still predictable and readable.
Ten tools is well-scoped for a wallet with policy enforcement: creation, inspection, policy, funding, pre-flight check, payment, approval flow, audit, and kill switch each earn their place. Nothing feels padded or missing at the count level.
The core lifecycle (create, policy, pay, approve/deny, audit, revoke) is well covered, but the approval workflow has a notable gap: there is no tool to list pending payment requests, so approve_payment/deny_payment require IDs an agent cannot discover. There is also no unfreeze/unrevoke or wallet-listing operation, which limits recovery and multi-wallet workflows.
Maintenance
Related MCP Connectors
Give AI agents a wallet with a spending limit. Non-custodial USDC on Base, server-enforced limits.
Pre-spend firewall for AI agents. Approves, blocks, flags transactions against policy rules.
Payment infrastructure for AI agents: spending rules, approval flows, single-use virtual cards.
Give your AI an allowance: pay approved people within on-chain spending limits.
Related MCP Servers
- AlicenseAqualityDmaintenanceEnables AI agents to check balances and send transactions across multiple blockchains with automatic spending limit protection and policy enforcement.3MIT
- FlicenseNot gradedqualityDmaintenanceEnables AI agents to propose wallet payments while a local, human-authored policy decides whether each transaction is approved, requires human confirmation, or is refused, and records every decision in a signed, append-only ledger.-
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to propose structured payment intents, apply deterministic policy, require human approval, and record tamper-evident audit trails without connecting to live payment rails.MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to request policy-governed payments through MCP tools, with server-side enforcement of spending limits, merchant allow-lists, Pix BR Code validation, approval thresholds, and hash-chained audit logging in a sandbox. It lets agents check wallets, create card or Pix payment intents, and poll or list payments while keeping approval and limit changes out of agent control.Apache 2.0