Comparing Harnesses for AI Trading on Moby

An always-on AI agent that scans Moby for new Solana tokens can cost $3 a month or $3,000. The difference is mostly the harness you run it in, not the model.

We ran Moby's MCP server through two agent harnesses, DeepSeek Harness and Hermes, and priced the same scanning workload across pay-per-token APIs and flat-rate plans. Here's what we learned, and what we'd pick at each budget.

What is an agent harness?

A harness is the agent program that runs a model against Moby's MCP server. A trading agent has three layers, and each one affects your bill:

• The harness is the agent program. It holds the prompt, calls Moby's tools, and decides how much history each request carries. DeepSeek Harness and Hermes are two examples.
• The model does the reasoning: DeepSeek V4.1 Flash, GLM 5.3 Flash, Qwen3.8 Flash and so on.
• The provider serves the model and bills you, either per token (OpenRouter, the DeepSeek API) or as a monthly plan with an allowance (OpenCode Go, OpenDesign).

Most comparisons focus on the model. For a loop that runs every few minutes, the harness matters more.

Why the harness matters most

The two harnesses we tested handle memory in opposite ways, and that sets the cost.

DeepSeek Harness keeps one long conversation. Every scan adds to it, so by the end of a two-hour session each request was re-reading about 200,000 tokens of history. A single scan used about 1.2 million tokens in total. It stays affordable only because DeepSeek bills cached, repeated tokens at a tiny $0.003 per million.

Hermes starts every scan fresh. Each scheduled run sends a short prompt plus the tool results it needs, then exits. A scan used 21,000 to 47,000 tokens: 26 to 58 times less than the long-context approach.

That gap decides which models you can afford. A long-context harness only makes sense on models with very cheap cached reads. A stateless harness like Hermes works well on almost any budget model and fits inside low-cost monthly plans.

What each setup costs

Estimated monthly cost of one scanning strategy running around the clock, at three cadences:

Ranges run from a light, mostly cached scan to a heavy one with nothing cached. Estimates from our own test sessions. Prices are list prices as of November 2026 and change often, so check each provider before you commit.

A few things stand out:

• GLM 5.3 Flash is the value pick. It's among the cheapest models on OpenRouter and scored highest on agentic benchmarks among the budget models we compared.
• $10 plans cover most cadences. OpenDesign Go buys $120 of DeepSeek V4.1 Flash usage a month and OpenCode Go allows $100, enough for any 5 to 10 minute DeepSeek loop.
• Premium models don't fit polling. Re-reading long context on every scan at frontier prices runs into thousands of dollars a month.

Picking a cadence: start at 5 minutes

Faster isn't always better. In our tests:

• A "1-minute" schedule didn't scan every minute. Each Hermes run took 2 to 4 minutes, so busy slots were skipped and the agent settled at about one scan every 3 minutes.
• Our strategy didn't need it. The Early Launch Hunter strategy confirms signals over 30 minutes, so 5-minute checks caught the same signals as 1-minute ones at a fifth of the cost.
• Flat-rate plans dislike it. A true 1-minute loop sends hundreds of thousands of requests a month, the kind of traffic that can trip a plan's fair-use monitoring.

Start at 5 or 10 minutes. Move faster only if your strategy's signals genuinely appear and fade within minutes.

Our picks

• One strategy, lowest cost: Hermes + GLM 5.3 Flash on OpenRouter. About $6 to $14 a month at 5 minutes, billed only for what you use. Set its reasoning effort to low for scans; it defaults to maximum.
• DeepSeek fans: Hermes + DeepSeek V4.1 Flash on OpenDesign Go. $10 a month ($7 billed annually) covers a 5 to 10 minute loop for the whole month, and Hermes is a listed supported agent. OpenCode Go at $10 is a close second.
• Several strategies at once: a $10 plan. OpenCode Go gives each model its own allowance, so a GLM strategy and a DeepSeek strategy at 10 minutes can share one subscription.
• Already on DeepSeek Harness: keep it on the DeepSeek API. Its long context relies on cheap cached reads, so stay on providers that charge around $0.003 to $0.005 per million cached tokens.

Spend fewer tokens with any harness

Whichever setup you choose, these habits cut the bill:

• Run one scheduled job per strategy. A duplicate job in one of our tests added 30 to 100% to usage.
• Give the agent only the tools it needs. Every tool definition is re-sent on every request; a focused set of scanner tools beats the full catalog.
• Use Moby's batch tools. get_token_trade_brief bundles pre-buy safety and smart-money checks into one call, and get_whale_wallet_profiles looks up to 20 wallets at once.
• Ask for less data. Keep feed limits at 10 to 20 rows instead of 100.
• Cap the output. Ask for a short structured verdict with a low token limit, and save deep reasoning for actual buy decisions.
• Keep the prompt identical between runs. A stable prompt at the start of each request lets providers bill it at cached rates.
• Measure. Log tokens per run from your provider's usage reports, and compare before and after each change.

Before your agent trades

A cheap harness still trades real money. On Moby:

• Your agent trades Solana tokens only, funded in USDC, from its own agent wallet. Each user has one agent.
• The agent cannot withdraw funds. You withdraw in the Moby app.
• Fees: A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees. Note that at a 5-minute cadence, a strategy that buys often pays this fee on every swap, on top of model costs.
• Agent buys have no automatic exits. If you want take-profit or stop-loss behavior, ask your agent for it explicitly.
• You set your agent's limits on max per trade, daily buy volume and market-cap range from the Moby app, and Moby checks them server-side on every buy your agent makes. Read AI agent guardrails and AI agent security before you leave a loop running.

Get started

Connect Moby's MCP server to Hermes, point Hermes at OpenRouter, OpenDesign or OpenCode Go, and start with a 5-minute schedule and a single strategy. Authenticate with your agent key as a Bearer header. Any MCP client that can send an Authorization header can connect using the generic setup.

Prefer a coding agent host? See the Codex setup, the Claude Code setup, or step-by-step agent tutorials. Browse strategies in the Moby marketplace and start at Moby agents or read about agentic trading on Moby.

Costs are estimates from our own test sessions and published list prices as of November 2026; your usage and prices will differ. Check each provider's terms, since some plans are built for coding agents and monitor other traffic. AI agents can make mistakes, and crypto trading carries risk. Nothing here is financial advice.

FAQ

What is an agent harness?

An agent harness is the program that runs an AI model against tools such as Moby's MCP server. It holds the prompt, calls the tools, and decides how much history each request carries. Hermes and DeepSeek Harness are two examples.

How much does an always-on AI trading agent cost to run?

In our tests, one scanning strategy on Moby cost about $6 to $14 a month at a 5-minute cadence with Hermes and GLM 5.3 Flash on OpenRouter, and $1,500 or more a month with a premium frontier model and long context. These are estimates from November 2026 list prices; check each provider before you commit. Moby also charges a 0.75% swap fee (minimum $0.90 per swap), plus onchain fees, on every trade.

Why does the harness matter more than the model for cost?

A long-context harness re-reads its whole history on every scan, so one scan used about 1.2 million tokens in our tests. A stateless harness like Hermes starts fresh each run and used 21,000 to 47,000 tokens per scan, 26 to 58 times less.

How often should a trading agent scan?

Start at every 5 or 10 minutes. In our tests, 5-minute checks caught the same signals as 1-minute ones for our strategy at a fifth of the cost, and true 1-minute loops can trip fair-use limits on flat-rate plans.

How do I connect a harness to Moby?

Authenticate with your agent key as a Bearer header. Any MCP client that can send an Authorization header can connect using the generic setup. Setup guides for Codex and Claude Code are on moby.win/agents.

Can my agent withdraw funds or exit trades on its own?

No. The agent cannot withdraw; you withdraw in the Moby app. Agent buys have no automatic exits unless you ask your agent to manage them.