Hooking a wallet to Claude Code: the authorization holes in wallet MCP servers
Objective_Mix2531 · reddit · 2026-10-14
A long Reddit post examines what happens when you connect a crypto wallet to Claude Code via wallet MCP: how do you ensure an agent that prepares a transaction can't submit one?
- Architecture: keep key import/backup outside the conversation, sign locally, never expose the seed phrase (examining iron wallet mcp)
- Initial eval: dedicated small-balance wallet in read-only mode; test whether the agent picks the right wallet/asset/chain under ambiguous requests
- Policy controls: server-wide read-only, per-tx USD cap, recipient allowlist — but verify they're actually enforced; unclear how allowlists apply to swaps
- Key hole: a policy-setting tool is in the tool list — can the agent change its own limits?
- Deeper risk: keeping the seed out of context doesn't stop an agent hitting malicious instructions and using an already-permitted signing tool, or simply misunderstanding you
- Approval semantics: does approval for an exact amount/recipient survive mid-execution changes? Cumulative budget? Stale USD prices? Retry behavior after timeouts?
The post doubles as a solid security-audit checklist for financial MCP tools.
More from coding & agent
- Developer shares full trace of local AI agent code review with an 'armed fence' approval workflow — HankYeomans · 2026-10-14
- Dev builds scope-locked local agent harness: GLM 5.3 Flash fixes a real bug PR-by-PR — HankYeomans · 2026-10-14
- Claude Code session startles dev with a familiar name — turns out it's MCP Agent Mail — doodlestein · 2026-10-14
- LSCBench: Vibe Coding Agents Write Functional Code, But Only 25% Is Also Secure — Qwen · 2026-10-14
- sshexport 2.9 ships with ChatGPT-audited hardened code and OpenWRT support — esrtweet · 2026-10-14
- AI agent deletes entire edited season of Swedish reality show weeks before premiere — torb · 2026-10-14