Agent Wallet

ChainGPT Agent Wallet Skill

The agent has its own EOA wallet on every EVM chain it supports. The admin (you, in your shell) sets policies that the agent cannot violate or revoke — even if a malicious prompt convinces the LLM to try.

Threat model

The attacker's goal: prompt-inject the agent to drain its wallet to an attacker address.

The plugin's defense: the policy check is in code, not in the LLM's prompt. Every chaingpt_agent_wallet_sign_and_send call:

1. Loads the policy file fresh from disk (no caching — admin can update mid-session).
2. Runs checkPolicy(intent) — pure deterministic code that doesn't see the LLM's context.
3. Refuses if any rule fails, with a clear reason the agent surfaces back to the user.

The attacker can convince the LLM to call sign_and_send(to=attacker, value=ALL) — but the tool layer refuses because attacker isn't in allowedToAddresses or value exceeds maxTxValueWei or killSwitch=true. The trust boundary is the tool code, not the LLM.

There is no MCP tool that writes the policy file. The admin edits it directly with a text editor. There is no MCP tool that reads or sets the passphrase. The passphrase lives only in the shell env var or the OS keychain — never in the keystore file, never in the LLM's context.

Setup (admin steps — done once)

The keystore passphrase resolves in this priority order:

1. CHAINGPT_AGENT_WALLET_PASSPHRASE env var — explicit override. Best for CI / headless / power users who want zero process-list and zero keychain exposure.
2. OS keychain — auto-managed. On macOS (Keychain via security) or Linux (libsecret via secret-tool), if no env var is set, chaingpt_agent_wallet_init generates a strong 256-bit passphrase and stores it in the keychain. You never type or remember it; the MCP server reads it back on each load.

Option A — zero-setup (macOS / Linux with a keychain) — recommended for most

# Just init — a strong passphrase is generated + stored in your OS keychain.
claude
> initialize the agent wallet

The init output tells you it used the keychain and how to export the passphrase for backup.

Option B — explicit env var (CI / headless / max control)

# Set a strong passphrase BEFORE starting the MCP server (>= 16 chars)
export CHAINGPT_AGENT_WALLET_PASSPHRASE="your-strong-passphrase-here-min-16-chars"
claude
> initialize the agent wallet

Back up the passphrase either way. Keychain entry: service=chaingpt-mcp-agent-wallet account=keystore-passphrase. Export on macOS with security find-generic-password -s chaingpt-mcp-agent-wallet -a keystore-passphrase -w. Lose it (and any backup) → the keystore is unrecoverable. There is no recovery path.

Security tradeoff of the keychain option: the secret stays out of plaintext-on-disk and out of the LLM context, but the keychain is unlocked while you're logged in — a local attacker on an unlocked session could read it. That's a much higher bar than a plaintext file and appropriate for a low-value bounded hot wallet. For zero local exposure, use Option B.

This creates two files:

Both default to ~/.chaingpt-mcp/agent-wallet/ but can be overridden via CHAINGPT_KEYSTORE_FILE and CHAINGPT_AGENT_POLICY_FILE.

Tools

Policy file format

Default policy.json (lazily created on first read) is the Balanced DeFi policy: killSwitch: false, major DEX/lending routers allow-listed, 0.1 native per-tx cap, 0.3 native + 20 txs per rolling 24h (maxDailySpendWei / maxDailyTxCount), memo required. A corrupt or partially-missing policy file always falls back to fail-closed (killSwitch: true) — tampering can never open the gates. Apply the "Locked down" template (or set killSwitch: true) for a refuse-everything posture.

Example production policy (allow DEX rebalancing on Base, capped at 0.1 ETH/tx, audit memo required):

{
  "version": 1,
  "killSwitch": false,
  "allowedChains": [8453],
  "allowedToAddresses": [
    "0x6352a56caadc4f1e25cd6c75970fa768a3304e64",
    "0x111111125421ca6dc452d289314280a0f8842a65"
  ],
  "blockedToAddresses": [
    "0x0000000000000000000000000000000000000000"
  ],
  "maxTxValueWei": "100000000000000000",
  "maxTxGas": "500000",
  "blockedSelectors": [],
  "requireMemo": true,
  "notes": "Base only, OpenOcean + 1inch routers only, 0.1 ETH cap, memo required for audit",
  "updatedAt": "2026-05-18T20:00:00Z"
}

Field reference

Precedence: kill switch > blockedToAddresses > allowedToAddresses > value caps > gas cap > daily velocity caps > blockedSelectors > memo. Any single failure refuses the tx.

Solana wallet (v1.19+)

The same bounded-autonomy model on Solana. Separate Ed25519 keystore (solana-keystore.json, same cipher + same admin passphrase), gated by the solana policy sub-object that — like everything else here — no MCP tool can write.

chaingpt_agent_wallet_solana_init           # one-time keystore
chaingpt_agent_wallet_solana_address        # fund this (the balance is the outermost cap)
<any builder: jupiter swap / marginfi / kamino / transfer>  → unsigned VersionedTransaction (base64)
chaingpt_agent_wallet_solana_sign_and_send txBase64=<…> memo=<…>

Policy block (admin-only, dashboard or text editor):

"solana": {
  "enabled": true,
  "allowedPrograms": ["11111111111111111111111111111111", "JUP6LkbZbjS1jKKwapdHNy74zcZ3tLUZoi5QNyVTaV4"],
  "maxTxLamports": "100000000",
  "maxDailySpendLamports": "300000000",
  "maxDailyTxCount": 20,
  "requireMemo": true
}

Hard facts to relay accurately:

• Fail-closed migration: a policy file without solana.enabled: true refuses every Solana signing op. unrestricted does not bypass it.
• Spend is simulation-priced: the tool simulates first and meters the fee-payer lamport delta against maxTxLamports + the 24h window. Simulation unavailable or failing ⇒ refusal, never a blind broadcast.
• The program allowlist fences which protocols the agent may ENTER (top-level instructions). Inner CPIs are invisible to it — the lamport caps + tx count are the actual spend fence. SPL-token outflows don't move fee-payer lamports, so they are fenced by the allowlist and tx count, not the lamport cap.
• The agent must be the sole signer and fee payer — co-signing or fee-sponsoring someone else's tx is refused structurally.

On-chain caps — ERC-4337 session keys (v1.21+)

The endgame: the user's ERC-7579 smart account grants the agent's EOA a SCOPED on-chain session (Smart Sessions module). The caps live in audited contracts and are validated by the EntryPoint. Status: BETA — the module addresses are verified deployed on Base Sepolia and the encoders are unit-tested, but the end-to-end live proof is not yet published. Treat the on-chain column below as the DESIGNED guarantee, not yet an independently-demonstrated one; the local gate is your tested fence today.

chaingpt_aa_session_build_grant chain=base account=<user SCW> tokenCaps=[{token: USDC, cap: "100000000"}] validUntil=<unix>
  → OWNER signs the userOpHash externally → chaingpt_aa_submit_userop
chaingpt_aa_session_status                    # chain-authoritative: enabled? remaining?
chaingpt_agent_wallet_4337_sign_and_send …    # the agent acts; local gates AND chain caps both apply
chaingpt_aa_session_build_revoke …            # incident response: chain-level kill

Hard facts: erc4337.enabled policy opt-in is fail-closed and OFF by default everywhere (this surface acts on a third-party account). Unbounded grants are refused at build time. A bundler rejection of an over-cap op is the product working — never retry around it. v1 supports Biconomy Nexus 1.x accounts.

Pre-flight checklist

1. chaingpt_agent_wallet_status   # see address + policy digest + kill switch state
2. chaingpt_agent_wallet_balances # confirm funded on the target chain
3. chaingpt_agent_wallet_policy   # read the active rules in full
4. chaingpt_agent_wallet_sign_and_send chain=… to=… valueWei=… data=… memo="…"

If the call gets refused with a policy reason: do not try to work around it from the agent side. Surface the reason to the admin and let them edit the policy file (or override) themselves.

Local admin dashboard

> Use chaingpt_agent_wallet_serve_ui

Returns a http://127.0.0.1:8787 URL and a one-time admin token printed in the tool output (also saved to ~/.chaingpt-mcp/agent-wallet/.admin-token, 0600). The token rotates on every restart.

Open the URL in your browser. Paste the admin token at the login screen. The dashboard then shows:

• Deposit address with QR code
• Multi-chain native-coin balances (refresh page to update)
• One-click kill switch — engage or disable, single button
• Policy JSON editor — full inline editor, validated server-side, atomic write with .bak backup
• Keystore + policy file paths for reference

Why this is safe even though the dashboard CAN edit the policy

Recall the threat model: the attacker controls the LLM via prompt injection. The defenses, in layers:

1. No MCP tool exposes a write to the policy file. The LLM literally cannot reach the savePolicy function — it's only imported by the localhost HTTP server.
2. The localhost HTTP server has no client inside the agent. The plugin has no MCP tool that does arbitrary HTTP POSTs to localhost. The LLM cannot trigger the dashboard's POST endpoints even by trying.
3. Admin auth required. Even if a future tool somehow gained HTTP access, every POST endpoint requires a valid session cookie that's only set after the admin pastes the token. The token rotates every restart and lives only in admin-controlled state (env / file with 0600 perms).
4. Origin + Referer check. Cross-origin POSTs (CSRF) are rejected. Browsers always set Origin on form submits.
5. Strict schema validation. Even with a valid session, the policy editor rejects unknown fields, bad chain IDs, malformed addresses, non-integer wei values, etc. Garbage cannot make it onto disk.
6. Atomic write + .bak. A botched save can't corrupt the policy file mid-write, and the previous version is recoverable.
7. Bound to 127.0.0.1 only. Never on 0.0.0.0 — the dashboard is not reachable from other machines on the network.

The single failure mode that would bypass all of this: malware running on the admin's machine that can read the admin token file AND make HTTP requests to localhost. At that point the attacker has shell access and can read the keystore directly; the policy file is no longer the weakest link.

Dashboard endpoints

What this skill does NOT do

• Solana / non-EVM signing. The agent wallet is EVM-only. Solana program-instruction signing is a separate path not yet wired.
• Multi-sig. This is a single EOA. For larger sums use a Safe / Gnosis multi-sig and have the agent's EOA be one signer; the policy still applies to the agent's signing.
• Session keys / ERC-4337. Bounded smart-account autonomy is a different model (smart account holds funds, agent gets a revocable session key). This skill is the simpler EOA + policy-file approach; ERC-4337 is a future feature.
• Recover lost passphrases. Lose CHAINGPT_AGENT_WALLET_PASSPHRASE → the keystore is unrecoverable. There is no recovery path. Back up the passphrase out-of-band (1Password, hardware safe, etc.).
• Read or change the policy from the agent's tools. Deliberately. Edit the file with your text editor.

Funding the agent

1. Get the address: chaingpt_agent_wallet_address
2. Send funds from any wallet (CEX withdrawal, MetaMask, hardware wallet, etc.) to that address on any chain the policy allows.
3. Confirm: chaingpt_agent_wallet_balances chains=[base,arbitrum]

The agent is now ready to operate within its policy bounds.

Credit accounting

All agent-wallet tools cost 0 ChainGPT credits. The wallet is custody-on-the-user's-machine (not custody-via-ChainGPT). The credit funnel comes from upstream tools the user/agent calls (chaingpt_research_token, chaingpt_risk_token, chaingpt_intel_token) before deciding to deploy.

Reference

• Keystore format: AES-256-GCM with scrypt (N=2^14) KDF. File version 1.
• Default paths: ~/.chaingpt-mcp/agent-wallet/{keystore.json,policy.json} (override via env).
• File perms: 0600 for keystore, 0700 for parent dir (POSIX).