AEREDIUM, a digital asset infrastructure company, announced two products on October 22 that check what an AI agent may do before it does it. The Chair supervises coding agents. MCP Connector governs agents that make payments and trades. Both run through the company’s new AI platform , and AEREDIUM says they went live on October 13. A policy engine answers every proposed action with one of three outcomes: approved, denied or held for a person.
The launch lands as agents move from drafting text to touching real systems. Coinbase opened payment rails for AI agents to corporate clients in July, and in April the founder of a car-rental software startup said a coding agent deleted his production database and its backups in one nine-second API call, Decrypt reported. OWASP maps prompt injection to six of the ten categories in its Top 10 for Agentic Applications, Help Net Security noted in June. AEREDIUM CEO Albert Dadon frames the problem simply: the challenge “isn’t just making these agents more capable, but making sure they operate within the rules we set.”
How The Chair supervises coding agents
The Chair sits between an AI coding agent and an organization’s development environment, and it checks each tool call before the call runs. AEREDIUM says that covers executing commands, modifying files, merging code and pushing changes. Calls the organization’s rules already cover are approved or denied automatically. Everything else is held. Held requests land in a control panel, where authorized users approve or reject them, and each verdict produces a signed receipt that records what the agent attempted, which policy applied and who authorized it.
The product page adds detail the announcement leaves out. Every command is read twice, the second time by a model that asks twelve questions, among them whether the command reaches the network, installs anything, deletes or starts a machine. Where that second reading can’t rule out an act the policy forbids, the receipt says so. A Claude Code plugin handles integration, and a governed shell covers other agents, including OpenAI’s. Teams build policy by answering a questionnaire, not by writing code. After a 30-day trial, The Chair costs $10 per person a month, or $80 a year.
MCP Connector puts a wallet behind the rules
MCP Connector pairs the same policy engine with a digital asset wallet. Each agent gets its own wallet, its own policy and a designated owner. The owner picks the agent’s role, approves its mandate with a passkey and can halt or delete the agent at any time. Two roles exist today. A Payer can send USDC, USDT, PYUSD and EURC to addresses the owner has approved, but it can’t trade. A Trader can trade on designated exchanges inside its own wallet, but it can’t send funds out. The product page names Curve, Uniswap and PancakeSwap as the venues.
Custody is the central claim. AEREDIUM says the wallet signs with threshold signatures generated inside secure enclaves, so no complete private key is ever created and the company holds no key to the owner’s funds. MCP Connector supports Ethereum, Arbitrum and Base and keeps transaction records for seven years. It’s designed for Claude, and the company says it’s open to OpenAI and other frameworks. Plans start at $49 a month for three agents. Trades carry a 0.05% service fee, and transfers carry none.
Who it is for, and what isn’t proven yet
The Chair targets software teams that use AI coding agents, and MCP Connector targets organizations that let agents handle money. Signed, retained records also give audit and compliance staff something to check. Both tools belong to a wider build-out of agent infrastructure. Binance’s Agent OS , live since August 20, lets outside AI tools read account data and place supported trades with the user’s permission, and XDC Network picked Concordium as the identity verification layer for its agent payments marketplace on October 14.
Plenty is still unproven. The release names no customers, and neither product page cites an outside audit of the signing design. The Chair’s second reading is a model’s judgment, and the page gives no accuracy figures. Held requests also depend on someone answering them. Demand is the bigger unknown. TRM Labs’ September analysis called x402 the most widely used agent payment rail, yet put the share of its screened commerce that appears to be agentic at between 0.6% and 7.5%. And MCP Connector enables just two of the policy engine’s twelve roles so far.
The next signals are concrete: named customers, an independent review of the custody design, and data on how often the engine holds or blocks an agent in live use.