hermes · tool-calling · 312 runs today
Run Hermes agents without building the runtime
Hermes models are good at calling tools. Turning that into something dependable means a loop, a tool registry, retries and somewhere to keep state — that is what we host.
- Bring your own endpoint
- Tool registry included
- Full run traces
From £14.99 a month, no minimum term. Prices exclude VAT. No setup fee.
every run, recorded
When an agent does something odd, you can see where it turned
Aggregate metrics tell you a run failed. A trace tells you which tool it reached for, what came back, and which step sent it sideways. Every run keeps one.
Need current pricing for SKU AX-4120 from three suppliers.
http_get(supplier_a.api/price?sku=AX-4120)
£184.00 · in stock · 3 day lead
sql_query(SELECT last_paid FROM purchases WHERE sku=…)
last paid £201.50 on 12 Aug
Supplier A is £17.50 cheaper than last purchase. Draft raised.
- agents on Pro
- 0agents on Pro
- trace retention on Scale
- 0 motrace retention on Scale
- network uptime
- 0.0%network uptime
- UK support
- 0/7UK support
the options
What you actually have to build
The model call is a few lines. Everything that makes an agent survive contact with a real API is the rest of it.
| compared on | alternativeBuild the loop yourself | alternativeA closed agent platform | recommendedThe Email Shop |
|---|---|---|---|
| Model choice | Anything you like | Whatever the platform supports | Any Hermes endpoint — Nous, OpenRouter, Together or your own |
| Tool calling | You write the registry, schemas and dispatch | Built in, within their catalogue | Registry included; register any HTTP or SQL tool you like |
| Retries and failures | You handle timeouts, malformed calls and loops | Handled, opaquely | Backoff, retry limits and loop detection, all visible to you |
| State between runs | A database you set up | Theirs, in their format | Postgres on your instance, that you can query directly |
| Seeing what happened | Whatever you logged | Their dashboard | Full traces — every thought, call, response and timing |
| Where inference happens | Your choice | Their infrastructure | Your choice, and your bill — we do not resell tokens |
Model choice
- Build the loop yourself
- Anything you like
- A closed agent platform
- Whatever the platform supports
- The Email Shop · recommended
- Any Hermes endpoint — Nous, OpenRouter, Together or your own
Tool calling
- Build the loop yourself
- You write the registry, schemas and dispatch
- A closed agent platform
- Built in, within their catalogue
- The Email Shop · recommended
- Registry included; register any HTTP or SQL tool you like
Retries and failures
- Build the loop yourself
- You handle timeouts, malformed calls and loops
- A closed agent platform
- Handled, opaquely
- The Email Shop · recommended
- Backoff, retry limits and loop detection, all visible to you
State between runs
- Build the loop yourself
- A database you set up
- A closed agent platform
- Theirs, in their format
- The Email Shop · recommended
- Postgres on your instance, that you can query directly
Seeing what happened
- Build the loop yourself
- Whatever you logged
- A closed agent platform
- Their dashboard
- The Email Shop · recommended
- Full traces — every thought, call, response and timing
Where inference happens
- Build the loop yourself
- Your choice
- A closed agent platform
- Their infrastructure
- The Email Shop · recommended
- Your choice, and your bill — we do not resell tokens
If you need dedicated GPU inference rather than an API endpoint, this is not that product — talk to us about dedicated servers instead.
The runtime, not the model
We host the machinery around the model. Inference stays wherever you choose to buy it.
The agent loop, supervised
Plan, call, observe, repeat — with iteration limits, timeouts and loop detection so a confused agent stops rather than spending your budget in a circle.
A tool registry you control
Register HTTP endpoints, SQL queries and shell commands as tools with typed schemas. Hermes handles the calling; you decide what exists to call.
Any Hermes endpoint
Point the runtime at Nous Research, OpenRouter, Together or your own self-hosted vLLM server. Swap endpoint without rewriting your agent.
State you can inspect
Conversation state, memory and run history live in a Postgres database on your instance. It is your data in ordinary tables, not a proprietary blob.
Traces for every run
Each run records its reasoning steps, tool arguments, responses, token counts and latency. When an agent does something odd you can see exactly where it turned.
Spend limits that bite
Cap tokens per run and per day. When a limit is reached the run stops and you are told, rather than finding out from an invoice at the end of the month.
Why the loop is the hard part
Hermes handles tool calling well. These are the four things that still go wrong once it is talking to real systems.
A malformed tool call should not end the run
Models occasionally emit arguments that do not match the schema — a string where a number belongs, a missing required field. Naive implementations either crash or pass the nonsense straight through to your API. The runtime validates every call against its schema, hands the model a specific error it can correct, and only gives up after a set number of attempts. Most malformed calls are fixed on the retry and you never hear about them.
Agents loop, and loops cost money
The classic failure is an agent that calls the same tool repeatedly because the response never quite satisfies it. Left alone overnight that is a real bill and nothing to show. We track repeated calls with identical arguments, stop the run when a threshold is crossed, and record why. Combined with per-run and per-day token caps, the worst case is a stopped run and a notification.
State has to outlive the process
An agent that forgets everything between runs cannot do anything that spans a week. Keeping state in memory works until the first restart. Every instance gets a Postgres database holding conversation state, memory and run history in ordinary tables — so it survives restarts, and you can query it with psql rather than through an API.
You need the trace before you need the dashboard
When an agent does something unexpected, aggregate metrics tell you it happened and nothing more. What you want is the individual run: what it was thinking, which tool it reached for, what came back, and where the reasoning went sideways. We keep the full trace for every run, so debugging starts from evidence rather than from a hypothesis.
pricing
Pricing that counts agents, not tokens
Inference is billed to you by your own provider. This is the runtime.
Prices are per month, billed monthly. No minimum term — cancel any month.
Starter
One agent, in production
2 vCPU · 4 GB RAM · 40 GB NVMe · 1 agent · 5 tools
- Supervised agent loop
- Tool registry with typed schemas
- Postgres state database
- +2 more
Pro
most chosenSeveral agents sharing tools
4 vCPU · 8 GB RAM · 100 GB NVMe · 5 agents · unlimited tools
- Everything in Starter
- Up to five agents on one instance
- Shared tool registry across agents
- +2 more
Scale
Agents doing continuous work
8 vCPU · 16 GB RAM · 200 GB NVMe · Unlimited agents
- Everything in Pro
- Unlimited agents and parallel runs
- Separate staging instance
- +2 more
Prices exclude VAT. No setup fee.
Frequently asked questions
Stop rebuilding the same agent loop
Bring an endpoint and a list of tools, and we will have your first agent running today. From £14.99 a month, no minimum term.