Mesh Helm Get early accessEarly access

One place to build, run and check your AI agents.

Test a job before it goes live. Catch the run that ended “ok” with nothing to show. Stop a loop while you sleep.

Mesh Helm is a Mac app that puts your agents, projects, models, tools and scheduled jobs in one window. Work with an agent in your code. Choose what each agent can use. Schedule the jobs. Then see what every run did, what it cost, and what needs you.

For macOS. Works with Hermes Agent, the free agent engine by Nous Research, and the setup installs it for you. Or it runs on Mesh, the engine inside Mesh Bot.

A sample setup: 4 agents, 3 models, 4 tool servers, 4 skills, 12 scheduled jobs.

The workspace, in five parts

Pick the folder and the agent: “in my-shop as coder”. You see each tool step as it happens and each file change as red and green lines. When the agent must ask, the question waits in the chat. “Undo files” puts the files back.

Recreated from the app, with sample data. The five buttons, the side list and the top strip all work.

A green tick is not a result.

This is the part that most agent tools leave to you. You have four agents and a dozen scheduled jobs. Every morning the job history is green. Then you find one of these three runs. Each one ended “ok”.

  1. Morning lead digest

    support, each day 06:00

    runs since Saturday3
    web_search401, each run
    output112 chars
    usual output2,840 chars

    result: ok

    Almost no result

    Since Saturday

    The lead digest sent nothing. The search tool answered “401” in every run. Each run still ended “ok”.

  2. Weekly customer update

    sales, 07:16

    toolsend_email
    recipients42
    first sent05:06
    sent again07:16, after a retry

    result: ok

    Sent twice: held

    Two times

    A job ran again after a retry. 42 customers got the same email twice.

    With the guard, the second send waits for your answer.

  3. Competitor price watch

    research, 01:35

    web_search41 calls
    the same stepagain and again
    cost$4.80 of $5.00

    result: ok

    Loop: 41 searches

    Before breakfast

    One research run made 41 web searches in a loop. Most of the day’s budget was gone by eight.

Hermes records all of this. The facts are in the run history, the logs and the usage page of each profile. The work is to find them and join them, one profile at a time, before a customer does. Mesh Helm does that work.

Your whole setup, in motion.

Four agents, twelve scheduled jobs, their tools, and the places they send to. Hermes Agent runs them. Mesh Helm stands in the middle and watches three places where a run goes wrong without an error.

  • A run on its way out
  • The guard: each outside action passes it
  • Something that needs you

A simulation with sample data. Scroll to fly in.

It asks before the same thing goes out twice.

An agent is about to send, post or create something. The same tool, the same target and the same content already went out in the last 7 days. The guard sends the call to the engine’s approval step.

sales, Weekly customer update: send_email to 42 recipients The job starts.

It never blocks by itself. You allow it or you skip it, and the ledger keeps your decision.

It doubts the run that ended “ok”.

Mesh Helm learns the usual output size of each scheduled job from its good runs. A run that ends without an error and comes back far below that size is flagged. It needs five good runs of a job first.

support, Morning lead digest: the last eight runs Eight runs. Eight green ticks.

Each job gets a score from 0 to 100 from its last 14 runs. You set the result rule for each job.

It stops a loop while you sleep.

The Night Brake runs inside the engine, so it works with the app closed. It looks at the last minutes of each profile: model calls, tokens, the same step again and again. Over your limit, that profile runs no tool until you release it.

research, Competitor price watch, 01:35
Web searches, this run0
Budget for today12 %
The same search, again and again. Nobody is awake.

It is off until you turn it on. You set the limits.

Four profiles. One list.

The mesh is a picture. This is the product. Needs you is one list for all your profiles, most urgent first. Each line has the likely cause, the facts behind it, and the safe next step.

Recreated from the app, with sample data.

Four profiles. One list.

  • The guard holds a repeat.

    The profile sales is about to send the weekly customer update to 42 recipients. The same tool, target and content went out 2 h 10 min ago. The guard sends the call to the approval step. It never blocks by itself: you allow it or you skip it.

  • The run check doubts an “ok”.

    The morning lead digest ended without an error, three times. Each run gave much less output than the usual run of this job. Likely cause: web_search answered “401”. The score went from 97 to 61.

  • A budget speaks up.

    The profile research is at 96 % of its budget for today: $4.80 of $5.00. Needs you shows a budget at 80 % and again over the limit. A budget does not stop anything by itself.

  • An agent waits for your answer.

    The profile coder asks to run git push --force-with-lease. The question waits in the same list as all the rest.

  • And an empty list says so.

    When you come back, Home shows what happened since you left, in five lines or fewer. If nothing needs you, the app tells you that too.

Mesh Helm
1 running 4 need you Automations $6.42 today▲ 3.1×

Needs you

Duplicate held salesOpen the session

Weekly customer update: send_email to 42 recipients

The same action already ran 2 h 10 min ago, from a session of the profile “sales”.

Automation supportRun againPause

Morning lead digest: the result is almost empty

Likely cause: the tool web_search failed with “401” in each of these runs. Score 61 (was 97).

Budget researchOpen Costs

research is at 96 % of its budget for today

$4.80 of $5.00. A budget does not stop anything by itself.

Waits for you coderOpen the session

Refactor billing webhooks

The agent asks to run: git push --force-with-lease

Recreated from the app, with sample data.

And the rest of your work is in the same window.

Three more pages of the app, as small models. The data is sample data. The buttons work.

Test a job with no send.

One click on a job starts a test run in a new session. Inside the engine, Mesh Helm refuses each outside action of that session: a message, a mail, a post.

  1. Read the customer list
  2. Write the email
  3. send_email to 42 recipients
Run it. Watch the email.

A trial run costs model money, as a real run does. It does not hold terminal and file tools, because the job needs them to do its work.

See what a change reaches.

The Map is built from the real config of each profile. Select a part.

Amber parts depend on it. Blue parts are what it uses. Before you turn off a skill or change a model, a preview names the automations the change reaches.

See where the money went.

$6.42

Today
3.1 × a usual day
This month
$118.40
Forecast
$273.20
  • research$4.80
  • coder$1.02
  • support$0.38
  • sales$0.22

Cost by profile, by model and by automation.

Also in the app.

Profiles and skills
Edit the soul, the memory, the model and the approval rule of each agent. The app keeps up to 60 old versions of each file.
MCP tool control
Each tool of each server: on or off, how often it was used, its last error. You get an alert when the text of a tool changes after you accepted it.
Security
What each agent can do outside, the names of its keys, and six kinds of risk. Each risk comes with its fix.
Changes
Versions of the setup of each agent. “It worked on Friday” gets an answer: the list of what is different now.
Steps
Each model call and tool step of a session, with its time and its share of the cost.
From Claude Code
Bring sessions, skills, MCP servers and deny rules over, item by item. Keys are left out.

See all 73 features, page by page.

Three working days.

The same window, for three kinds of work.

Monday 08:40

You run the automations.

Four lines in Needs you. A repeat send is held: you skip it. The lead digest is almost empty, with a likely cause: “401” from the search tool. You fix the key and run the job again. Ten minutes, not a round through four profiles.

Before a change

You design the system.

You want a cheaper model for the support profile. The preview names two automations that the change reaches. You save and run the test it offers. Then Security shows one profile that can send email with approvals off. You fix that too.

After a slow run

You build with agents.

The Steps tab shows 13 seconds, three model calls and two failed tool steps. The slowest step and the costly call are marked. No trace export, no account, and nothing left your Mac.

Why pay, when Hermes Desktop is free?

Hermes Desktop is free and it is good. It gives you chat, projects, profiles, scheduled jobs with run history, failure alerts, usage by model, MCP and skills. Keep it if you like it. You pay for Mesh Helm when your agents do real work without you, and you want to know what they did.

AreaHermes Desktop (free)Mesh Helm adds
Daily workChat with split panes, projects, Git review, and its own screens to set up models and channels.In the app: a job form with every option, a skills and MCP catalog, a task board, Git, logs and chat export. For models and channels, it opens the setup of Hermes in a Terminal window.
Scheduled jobsCreate, edit, run history, failure alerts.A check on each result. It flags a run that ends “ok” with almost no output. A score from 0 to 100 for each job. A job form that sets the result rule and the budget with the job.
Outside actionsApproval modes and allow lists.A question before an exact repeat of a send, post or create, from any profile.
What needs youLists for each feature and each profile.One ranked list across profiles, with the likely cause and the safe steps.
CostUsage by day and by model.Cost by profile and by automation. Budgets that tell you at 80 % and over the limit. A job budget that can stop the run and pause the job.
ChangesA warning on a risky model switch.A list of the automations a change reaches, before you save. A trial run of a job with no send, on other models too. Versions of the setup of each profile.
Runaway workThe engine can pause all new work with one command.A brake inside the engine that stops one profile in a loop, with the app closed. A line before a run that will probably fail.

Read the full comparison, with what Mesh Helm does not do today.

When you do not need Mesh Helm: you run one profile, you have no scheduled jobs, and your agents send nothing outside. Then Hermes Desktop is all you need.

It works with what you have.

You need it

Hermes Agent

The free, open-source engine by Nous Research that runs your agents. You install it yourself. Mesh Helm reads the Hermes folder you already have: your profiles, sessions, skills and scheduled jobs show at the first start.

Use both, or one

Hermes Desktop

Both apps read the same folder, and sessions from both show in one list. The checks run inside the engine, so they also work on a run that you start in Hermes Desktop or in the terminal.

Bring it with you

Claude Code

Mesh Helm finds your Claude Code projects and lists what carries over. You pick the sessions to continue. One button opens a project in your own Claude Code again. Claude models run in Mesh Helm with an Anthropic API key.

Hermes Agent and Hermes Desktop are free and open source, by Nous Research. Mesh Helm is a separate, independent app that works with them. The price is for Mesh Helm. Hermes Agent stays free.

Built to be trusted.

You give this app a view of every agent you run. So here is what it does, and what it does not do.

All eight points, and how the app shows its own health.

  • Local.

    No telemetry and no account. Your data stays on your Mac: in your Hermes folder (or in Mesh's folder, on the Mesh engine) and in the Mesh Helm folder. The app makes one network call of its own: it reads a small file on this website to see if a new version is there. It sends nothing about you, and you can turn it off.

  • Names, not keys.

    It shows which keys a profile uses, by name. It never reads a value.

  • The guard asks. It does not block.

    It sends a repeat to the approval step and waits. Two things refuse a tool call, and you start each one: the Night Brake, which is off until you turn it on, and a trial run. One empty file turns all of it off.

  • You can leave.

    Your sessions, profiles, skills and jobs stay in Hermes and work without Mesh Helm. Mesh Helm’s own data exports as one file.

Mesh Helm Pro: $39 a month, or $390 a year.

Try it free for 14 days, with no key and no card. Sales open soon; people on the early-access list hear first.

One thing does not change: you pay your model provider directly, as you do today. The checks and the Watchdog use no model tokens.

Questions.

Short answers. Each one is true for the app as it is today.

All 18 questions.

Does it replace Hermes Desktop?

It can, and it does not have to. Mesh Helm is its own window, with the chat, your projects, profiles, skills and scheduled jobs. Hermes Desktop keeps working beside it: both read the same Hermes folder, and sessions from both show in one list. For the less common model providers and channels, Mesh Helm opens the setup of Hermes in a Terminal window.

Does it work with my existing setup?

Yes. It reads the same Hermes folder. Your profiles, sessions, skills and scheduled jobs show at the first start. A setup page checks the engine, the plugin, the model and the login of each profile.

Is it only a nicer window?

No. The checks run inside the engine, as a small plugin. So the Duplicate Guard sees a send from any profile, chat or scheduled job that goes through Hermes tools, and the Night Brake works with the app closed. The window is where you read the result and do your daily work.

What happens when Hermes updates?

Hermes ships often. When the engine changes, Mesh Helm runs its checks one time. If a part changed, it pauses the feature that needs it and says so. The rest keeps working, and the chat is never paused by a check. After the update it counts the jobs that still pass their check. We do not promise that a new Hermes release never breaks a feature.

Does my code or my data leave my Mac?

Not through Mesh Helm. The app has no telemetry and no account. Its one network call reads a small file on this website to see if a new version is there; it sends nothing about you, and you can turn it off. Chat text goes to the model provider you chose, through Hermes, as before.

Do I pay for AI models as well?

Yes, to your provider, as you do now. The checks and the Watchdog use no model tokens.

Bring your agents into one window.

For macOS. Your agents, your data and your model keys stay on your computer.