Solaris AI MCP

Solaris AI MCP lets the AI agent you already use work inside your Solaris AI account. From its chat, the agent can look up nodes and tokens, build a workflow, dry-run it against live data, save it to your workflows, run it, and read the results. Anything that would trade in your wallet stays off until you approve it in Solaris AI.

It speaks the Model Context Protocol, so any agent that supports remote MCP servers with OAuth can connect: Claude, ChatGPT, Claude Code and Codex are covered below.

Connect an agent

The server URL is:

https://flowactions.solarisai.io/mcp

The same URL, with a copy button and the steps below, is on the Connections page under Connect an agent. Whichever agent you use, connecting opens a Solaris AI page where you sign in and choose what the agent may do. The agent never sees your password or your wallet keys: it gets a token that only carries the permissions you picked.

Claude

1. In Claude, open Settings, then Connectors. On some plans this is under Customize.
2. Choose Add custom connector, name it Solaris AI and paste the URL. Leave the OAuth client fields empty: Solaris AI registers the connector on its own.
3. Select Connect, sign in to Solaris AI on the page that opens, choose the permissions and select Connect again.

On Team and Enterprise plans an owner adds the connector for the organization first; members then connect it to their own Solaris AI account.

ChatGPT

1. In ChatGPT, open Settings, then Apps and Connectors, then Advanced settings, and turn on Developer mode. It is available on the plans OpenAI offers it on.
2. Create a connector named Solaris AI with the URL above and OAuth authentication.
3. Select Connect and choose the permissions on the Solaris AI page. In a chat, turn Solaris AI on from the tools menu.

Menu names in both chat apps change with plans and releases. If a label differs, look for custom connectors or developer mode in the same settings area.

Claude Code

claude mcp add --transport http solaris-ai https://flowactions.solarisai.io/mcp

Then run /mcp inside Claude Code and choose Authenticate for solaris-ai. Your browser opens the Solaris AI permissions page.

Codex

codex mcp add solaris-ai --url https://flowactions.solarisai.io/mcp
codex mcp login solaris-ai

codex mcp login opens the Solaris AI permissions page in your browser.

Permissions

The permissions page lists what the agent asked for, each with a switch. Turn off anything you do not want it to have; at least one must stay on. What each one allows:

See your account (account:read). Your wallet address, your credit balance, the names and types of the credentials you connected (never their secrets) and the apps you run. Agents use this to pick a credential you already have instead of asking you for keys.

Read workflows (workflow:read). The workflows you make available to agents (see below), page by page, with secrets and account-specific values replaced by { "$redacted": true }. It also covers the node catalog, each node's settings schema, the building guide, the recipe list and the app catalog.

Create and edit workflows (workflow:write). Save new workflows and change the ones it may edit. Every save goes through the same checks as the editor plus the money rule described below. An edit is made against the version the agent read: if you changed the workflow in the meantime, the edit is refused and the agent has to read it again.

Dry-run and run workflows (workflow:execute). Rehearse a workflow, which runs it with simulated swaps and returns samples of what each step would produce; start a live run of a workflow that cannot move money; and turn a workflow's schedule on or off. A workflow that can move money is never run live or scheduled through this permission.

Read runs (execution:read). The status, timing, per-step results and outputs of runs of the workflows it can see, with the same redaction as workflows.

Market data (data:read). Token search, prices, swap quotes and FOMO Terminal data: who to follow, the herd, traders, tokens, search and the leaderboard. Third-party text in these results (token names, handles) is passed to the agent as data, never as instructions.

Trades, with your approval (money:propose). Off on every connection until you turn it on. The agent can propose a swap, starting an app plan, and running or scheduling a workflow that swaps in your wallet. Each proposal waits for you to approve it in Solaris AI, unless it is a swap under the automatic approval amount you set. Turning it on asks for a daily limit and that amount, explained under Approvals.

Which workflows an agent can use

An agent does not see all of your workflows by default. It can use:

• Workflows it created. It can read, run and edit them. You can change that like any other workflow.
• Workflows you share. In Connections, under Workflows agents can use, or from the editor's title menu with Available to agents, choose Read, Read and run, or Read, run and edit for each workflow. App plans can be shared to read only.
• Everything, if you turn on Expose all my workflows to agents. Every workflow is then readable and runnable (app plans read only), except the ones you set to Not shared.

These choices apply to every agent you connected.

What agents cannot do

Send funds to another address. Transfers, payments to other wallets, token launches, bridges and any operation that can move funds out of your wallet are refused when an agent tries to save them, so an agent cannot build such a workflow at all. You can still build those yourself on the canvas.

Trade without you. A workflow that can swap is saved with its automatic triggers off. Agents cannot run it live or turn its schedule on; with trade proposals turned on they can only ask you, and you approve in Solaris AI.

See your secrets. API keys, webhook secrets, credentials and values that identify your accounts are hidden from agents in workflows, runs and errors, and secrets inside URLs (passwords and key parameters) are removed from every result. Secrets are recognized by field and by name. A key typed into free text, such as a Code step, is not recognized, so keep keys in credentials.

Edit around values it never saw. A workflow that holds hidden values this connection never supplied (for example a secret you typed on the canvas) can be read by the agent with those values hidden, but not edited by it: edit it on the canvas. In workflows it may edit, the agent sends hidden markers back unchanged and the stored values stay as they are.

Approvals

With trade proposals on, an agent has four ways to ask:

• Swap (propose_swap): buy a token with SOL, or sell a token for SOL, once, on Relay from your own wallet. The amount is checked and priced like an app plan: your balance, the minimum order size and the token's listing on Relay. The approval shows the token's address, the value at today's price, the output a quote expects right now, the price impact and the Solaris AI fee at your rate. The order is placed within two minutes of your approval or not at all, it is refused if the price impact at that moment is over 5%, and it is stopped if it has not filled two minutes after that. It is never retried, and only one agent sale of the same token runs at a time.
• Start an app plan (propose_start_app), for apps marked startable in the app catalog, such as DCA. The values are checked and priced exactly as the app's own Start does, and the plan must be sized in an amount or in USD, not as a percent of a balance, so it has a value to count against your limit.
• Run a workflow live, once (propose_live_run), for a workflow it may run that swaps in your wallet and has an active manual trigger.
• Turn on a workflow's automation (propose_schedule), which switches on its schedule, webhook or wallet-activity triggers.

Workflows with an operation that can send funds elsewhere are refused, multi-order plans have to go through the DCA app, and workflows that repeat steps in a ForEach loop are refused too: the number of items is only known when the run happens, so an approval could not say how many trades it allows. Before a proposal is made, your wallet must already authorize Solaris AI to sign; if it does not, the agent gets a link to your wallet page instead.

The agent gives you a link to an approval page. Open it signed in to Solaris AI as yourself; an agent cannot approve. The page shows:

• what will happen, in Solaris AI's words, never the agent's;
• every step that can run, with every setting of each step as stored: amounts, tokens, slippage and the rest, including steps that are switched off (marked as such). For a schedule it also shows the triggers that will be switched on and the cadence they will run at;
• for an app plan, its value at the current price and how much of today's limit you have already approved.

Approve or decline. A request is valid for 15 minutes (a swap for 5, so you approve it close to the price it was valued at), and a newer request of the same kind from the same agent replaces the older one.

An approval is bound to exactly what you saw. If the workflow changes after the agent proposed it, approving does nothing and the agent has to propose again. One approval starts at most one run or one plan, however many times it is clicked or retried.

The daily limit

Each connection with trade proposals has a daily limit on the swaps and app plans you approve from it: $100 by default, up to $10,000, counted per UTC day. You set it when you connect and can change it in Connections next to the agent. An approval that would go over it is refused until you raise the limit or the day turns. A swap or plan that is never placed gives its value back.

Live runs and schedules of workflows have no fixed value, since a workflow can compute its amounts as it runs. They do not count toward the limit; instead you approve each one on its own, with every setting in front of you.

On top of the limit, agent swaps are capped at 30 an hour and 200 a day per account.

Automatic approval for small swaps

Next to the daily limit you can set an amount, up to $250, under which a swap goes through without asking you. It starts at $0, which means every swap waits for you. A swap at or under it is approved the moment the agent proposes it, as long as it fits the daily limit and the swap caps; anything else waits for you as usual. The approval page and the card in the chat say when a swap was approved automatically. App plans, live runs and schedules always ask.

Cards in the chat

In agents that render MCP Apps, some results show as cards under the agent's message:

• Workflow preview, after the agent saves or compiles a workflow: a read-only canvas and Open in Solaris AI.
• Approval request, after a proposal: the summary, the steps and Review in Solaris AI. The card updates as you approve or decline, and for a swap follows the order until it fills or is stopped, but it never approves anything itself.
• Run receipt, after a run starts: the status of each step as the run progresses.

Agents that do not render cards get the same information as text.

Disconnect an agent

In Connections, under Connected agents, each agent shows what it may do and when it was last used. Disconnect stops every token of that connection at once and cancels its pending requests. To reconnect, start again from the agent.

An agent's sign-in lasts an hour and it renews it on its own. Each renewal is valid for 30 days, so a connection left unused for 30 days has to sign in again. A renewal used twice, which happens when one is copied, ends the whole connection.

Limits

To keep runs fast for everyone, each connection can make up to 120 requests a minute, each account up to 5,000 a day, 30 rehearsals an hour and 10 swap quotes a minute. An agent that goes over is told when to retry.