Agents · MCP

Your agent suggests. You sign.

Connect Claude, ChatGPT or your own bot to Conductor. It reads your automations and suggests changes. Conductor rebuilds and simulates every suggestion, and nothing happens on chain until your wallet signs it.

Open your automations →Agent docs →

Conductor's manager

Conductor's manager watches each automation, explains every run from on-chain facts, and suggests changes as transactions you review. It never signs, holds keys or moves funds.

It reads: runs, balances, caps, the fee budget, the pool and hook, claimable pons fees and operator health.

Findings are public, written from public chain data. Suggestions are visible to the owner only.
Every number comes from code. The wording comes from set templates.
A suggestion goes through the same checks, ceilings and simulation as a connected agent's.
Owners can pause it per automation. If the manager is off, automations keep running.

Manager docs →

Connect in 3 steps

01Create a keyOpen your automation, find Connected agents, pick what the agent may do and sign. The key is shown once.
02Add Conductor to your agentPoint it at the MCP URL with the key. Claude Code, Claude Desktop and any MCP client work today.
03Review and signSuggestions land on the automation page and in your inbox with the change, the checks and the simulation. Sign the ones you want.

What an agent can't do

Sign or send a transaction. Only your wallet signs.
Withdraw funds or move them anywhere.
Change the owner, add a recipient or add a target.
Unpause, turn a task on, or shorten a delay or an interval.
Hand you calldata. Conductor builds every transaction from its own mirror.
Go past its ceilings: at most 2× a cap, at most 0.1 ETH a day, at most 0.05 ETH of gas money.

MCP URL

https://useconductor.xyz/mcp

Stateless HTTP, JSON responses. With no key it serves whoami and three public reads: list_automations, get_automation and list_runs. simulate_task and proposals need a key.

shell

claude mcp add --transport http conductor https://useconductor.xyz/mcp --header "Authorization: Bearer cagt_<your key>"

Run it once in a terminal, then ask Claude about this automation.

Scopes

Scope	What it allows
read	Read only. Reads this automation's runs, policy, tasks and balances. It can't suggest anything.
propose:limited	Suggest safer changes. Reads, and can suggest: pause, turn a task off, revoke an operator, lower caps, cancel a queued rule change. You sign each one.
propose	Suggest any allowed change. Everything above, and can also suggest: raise caps (at most 2×, at most 0.1 ETH a day), deposit gas money (at most 0.05 ETH), lengthen a task's interval, apply a queued change. You sign each one.

Tools

whoamiThe key's name, owner, automations, scope and expiry.
list_automationsAutomations; with a key, the ones it was granted.
get_automationOne automation: policy, tasks, balances, operator, recent state.
list_runsRecent runs with facts Conductor reads from chain.
simulate_taskSimulate a task at the chain head. Results only, no calldata. Needs a key.
create_proposalSuggest one change. Returns its id, status, review link and expiry. Needs a propose scope.
get_proposalA suggestion's status and the change Conductor rendered from it.

ChatGPT and claude.ai connectors

ChatGPT connects without a key today, so it reads public data only and can't suggest changes. That changes when Conductor supports OAuth.

Agent docs →