Mesh Helm Get early accessEarly access

Everything in the app.

The home page tells the short story. This page has the details: what each check looks at, what it does, what you set yourself, and the full list of 73 features.

Six questions. The app answers each one.

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. The questions, the side list and the top strip all work. Try Activity and Settings in the side list too.

Seven kinds of line can show in Needs you. The most serious come first.

  • Duplicate held: a repeat send waits for your answer
  • Automation: a job failed, did not deliver, or came back almost empty
  • Connection: a channel or a part of the engine is down
  • Waits for you: an approval or a question
  • Spend: today is 3 times a usual day, or over your daily amount
  • Budget: at 80 % and over the limit
  • No output: a session is silent for 15 minutes in the middle of a turn

The three checks, in full.

What each check looks at, what it does, and what you set yourself.

Duplicate Guard

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, from any profile, chat or scheduled job. The guard sends the call to the engine’s approval step.

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

At the start it watches messages, emails, drafts, posts, new issues, new events and new records. You can change that list and the 7 days.

Run checks

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.

Each run gets one of six results: ok, failed, did not run, not delivered, almost no result, a tool failed. Each job gets a score from 0 to 100 from its last 14 runs.

You set the result rule for each job. Must the result reach its target? Is “found nothing” a good result? How small is too small?

The digest

It tells you when nothing needs you.

When you come back, Home shows what happened since you left, in five lines or fewer. If the list is empty, it says so.

With the app closed, the Watchdog can send the same alerts to a channel you connected to Hermes: Telegram, Slack, Discord, Signal, WhatsApp, email and six more. On Mesh, it sends to a bot's Telegram chat, also when Mesh itself has stopped. It uses no model tokens. It is off until you turn it on.

You choose how often it looks, from every 5 to every 60 minutes. It sends each problem once, and at most 4 messages in an hour. A test message shows that it works.

Test the change before it goes live.

A new model or a new prompt can break a job that worked for months. In Mesh Helm you try the change first, with nothing sent. Later you see what changed, and when.

Trial run

Run the 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, a new job, a memory note. Add up to two other models, or a changed task. One table shows which run passes the result rule, what each one cost and how long it took.

Changes

See what changed, and when.

The Changes page keeps versions of the setup of each profile: settings, soul, memory, skills and jobs. “It worked on Friday” gets an answer: the list of what is different now, with key values hidden. You can also put two runs of a job side by side.

After an engine update

Know what still works.

When Hermes changes, Mesh Helm runs its checks one time and pauses only the part that broke, with the reason. Then it watches the real runs: “5 of 5 jobs that ran passed their check”, or the names of the jobs that worked before and fail now.

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. The app tells you when a job used them before.

For the hours when nobody looks.

A loop can spend for hours at night. A job can fail at 06:00 for a reason that was in plain view at 05:30. These checks work before the damage, and they keep the proof after.

Night Brake

It stops a loop.

Off until you turn it on. It 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, many new sessions. Over your limit, that profile runs no tool until you release it, and its scheduled run is cancelled and paused.

Preflight

It looks ahead of a run.

In the 30 minutes before a job starts, Mesh Helm looks at what the job needs. Its channel is not connected, its profile is stopped, or its last runs all had the same problem: you get a line in Needs you before the run, not after.

Review and proof

It keeps the receipts.

Each finished run gets a receipt of facts: time, steps, what failed, what went out, what it cost. Accept it or send it back. Home counts what the checks caught in 30 days, and one click makes a report that you can give to a client.

See which profile can do damage.

The Security page has one block for each profile. It shows the approval rule, what the profile can do on this computer, what it can do outside, and the names of the keys it uses.

It marks six kinds of risk, the most serious first. Each one comes with the way to fix it.

  • Approvals are off and the profile can act.
  • A key is written in plain text in the config file.
  • A connected channel has no list of allowed users.
  • A channel, a local terminal and loose approvals in one profile.
  • The Duplicate Guard does not run in a profile that acts outside.
  • File or terminal tools with checkpoints off.

Mesh Helm reads the names of keys. It never reads or shows a value.

Each tool of each MCP server.

An MCP server gives your agent tools that you did not write. In the app, “Tools” on a server lists each one: on or off, what it can do, how often it was used, its last error. Turn off what you do not need.

A server can change the text of a tool after you started to trust it. Mesh Helm keeps what you accepted. When a description or an input changes, or a new tool comes, Home tells you, with the old text next to the new one.

Where the time and the money went.

Today, this month, the forecast for the month and a usual day. Cost by profile, by model and by automation. For each job: what one useful run costs. Budgets for a profile or a job.

Every step, with its time and cost.

The Steps tab lists each model call and tool step of a session, with its time. Each model call shows its share of the cost. The slowest step and the most costly call are marked, and subagents are listed. The recorder keeps facts only: no text of a tool call or a result.

The same error, read once.

Mesh Helm groups the same error across runs, tools and channels.

Needs you shows a budget at 80 % and again over the limit. For a scheduled job, you can turn on “Stop at the limit” in the app: Mesh Helm then stops the run and pauses the job. The app must be open for that, and the stop can come up to 1 minute late. A budget of a profile only tells you. The Limits page adds a monthly cap for each profile, as a bar that turns amber at 80 %.

Bring your Claude Code work with you.

  • Mesh Helm finds your Claude Code projects. It shows names and counts only, and it changes nothing.
  • Before you import, “What carries over” lists it. A note warns when Claude Code will delete a session soon.
  • You pick the sessions to continue. A session is never copied twice, and each copy can be removed again.
  • Skills, MCP servers and deny rules come over item by item. Keys are left out. Allow rules and hooks are never imported.
  • A brief tells the new session where the work stood. It is built from facts, with no model call, and you can edit it.
  • “Polish with AI” is your choice. It shows the token count and the cost first.
  • One button opens the project in your own Claude Code again.

Claude models run in Mesh Helm with an Anthropic API key.

Everything in the app.

73 features in six groups. The app reads the Hermes setup you already have: your profiles, sessions, skills and scheduled jobs show at the first start.

Work with an agent

18 features
Chat first
A large prompt box. Streaming answers, the agent’s thoughts, and a card for each tool step with its input and output. Stop at any time.
Pick the project and the profile
“in my-shop as coder”: the folder the agent works in, and the agent that does the work.
Model for one session
Change the model of the open session without a change to the profile.
Context meter
See how full the memory of the chat is. One click makes it smaller.
Pictures and files
Add a picture or a file to a message. A picture goes to the model as a picture.
Commands and the plan
Type / to see the commands of the engine. When the agent makes a plan, the app shows it with its progress.
Approvals in the chat
When an agent must ask, the question shows in the chat with the engine’s own choices.
Sessions
Rename, pin, archive, delete. Group by project or profile. Sort by time or name. Filter by profile.
One list for all sessions
Sessions from this app, the terminal, Hermes Desktop and your automations show together.
Projects
Make a project from a folder, rename it, hide it, open it in Finder.
Files panel
The files a session used, the full tree, open a file or copy its path.
Changes, Terminal, Preview
Each change as red and green lines. The output of the agent’s commands. A text preview of a file.
Undo files
The checkpoints of a session with their reason and files. See the changes, then put the files back.
Hand off
Go on as a different profile. The last 10 messages go with you.
Saved prompts
The text you use often, one click away.
Search
Titles at once, the text of all sessions from three letters, with the line that matched.
Command palette
⌘K: every page, session, automation and problem, and actions such as “New session as research” or “Set a budget”.
Keys
⌘K command list, ⌘N new session, ⌘F search, Esc to close.

Keep watch

18 features
Needs you
One ranked list across profiles, seven kinds of line, with the likely cause and the safe steps.
Guided recovery
A hint for the common errors (401, 429, timeout, 404, network). Run again or Pause from the line itself.
Run checks
Eight results for each scheduled run. Flags a run that ends “ok” with far less output than usual.
Result rule
You say what a good run gives, for each job: it must arrive, it must be new each time, it must have a number of items, the date of the day, or words that you name. A failed tool step can make the run bad.
Trial run
A test run of a job with no send. Compare the job as it is now with other models or a changed task: result rule, cost, time.
Preflight
A job that starts in the next 30 minutes and will probably fail gets its line before the run.
Night Brake
Stops a profile that goes in a loop or spends far too fast. Inside the engine, so it works with the app closed. Off until you turn it on. “Stop all” stops all new work.
Review
A receipt of facts for each finished run of the last 7 days. Accept it, or send it back with a note.
Monthly proof
What the checks caught in the last 30 days, on Home. A report as text, for you or for a client.
Score for each job
0 to 100 from the last 14 runs, with “3 bad runs in a row” and the next run time.
Duplicate Guard
Asks before an exact repeat of a send, post or create. You set the tools and the days. In the app, it can also ask for a message that is almost the same.
Repeated errors
The same error across runs, tools and channels, as one line.
Since you left
What happened while you were away, in five lines or fewer.
Top strip
Always in view: what runs, what needs you, the cost today, with a mark when today is far above a usual day.
Activity drawer
Running, Needs you, Automations and Spend, beside the chat.
Watchdog
Alerts to one of 12 kinds of channel with the app closed. No model tokens. Off until you turn it on.
Mac notifications
Five switches: a job fails, an always-on job is down, an agent waits, spend passes your daily amount, a turn finished. A test alert.
System health
For each profile: engine, sessions, costs, profile files, ok or broken, with the cause.

Understand the system

11 features
Map
Profiles, models, MCP servers, tool sets, skills, automations, channels and plugins, with what depends on each.
Change preview
The automations that a skill or model change reaches, before you save. A test run after.
Security
For each profile: outside powers, key names, six kinds of risk, each with its fix. In the app, each MCP server has a risk grade and a Test button.
MCP tool control
Each tool of each MCP server: on or off, what it can do, its use in 30 days, its last error. An alert when the text of a tool changes after you accepted it.
Changes
Versions of the setup of each profile: settings, soul, memory, skills and jobs. What changed and when, with key values hidden.
Compare two runs
Two runs of a job side by side: steps, time, cost, what failed.
Context X-ray
What fills the fixed part of each prompt: each tool set and each skill, in tokens, with an estimate for the month.
Steps
The time of each step and the cost share of each model call. Slowest step, most costly call, subagents.
Activity
The newest 100 sessions of all profiles: where each came from, its messages and its cost.
AI providers
The provider and model of each profile for main work, subagents, fallback and scheduled jobs. A warning when a login is missing.
Add a key the safe way
For Claude models: Terminal opens and the engine asks for your Anthropic API key. Mesh Helm never sees it.

Watch the spend

7 features
Costs
Today, this month, the month forecast, a usual day. By day, profile and model, for 7, 30 or 90 days.
Cost by automation
What each job cost, and what one useful run costs. Billed and estimate apart.
Budgets
For a profile or a job, each day or each month. A notice at 80 % and over. In the app, a job budget can also stop the run and pause the job.
Limits
A monthly cap for each profile, as a bar. A warning only.
Daily amount
One alert on the day your total spend passes it.
Spend today
Today against a usual day, 14 days of bars, and today’s most costly sessions. Home shows the spend of each profile.
Claude Code use
Your prompts in the last 5 hours and 7 days, next to the limit you enter.

Manage what runs

8 features
Profiles
Edit the soul, the memory and the About you file. Set the model, the thinking level, the approval rule and checkpoints. In the app: make a copy of a profile, compare two, export, import and delete. A delete makes an export first.
History
Up to 60 old versions of each profile file. Compare two, go back to one. A save is refused if the agent changed the file in between.
Test this profile
Start a session as the profile, straight from its page.
Skills
Search, turn on or off, edit with a backup, open in your editor, copy to other profiles. A removed skill goes to an archive.
Automations
The engine’s scheduled jobs: pause, resume or run. You type the name first. In the app, you also make or change a job, with its result rule and its budget in the same form. A new job starts only after its rule is saved.
Mac jobs
The background jobs on your Mac in plain words: what, why and how. Run, stop, archive or restore one. Jobs the engine needs are marked “do not stop”.
Find the session
See which chat made a Mac job.
Claude scheduled tasks
Listed beside your other jobs, to read.

Start and stay safe

11 features
Setup and help
Five checks at the first start: engine, plugin, model, login, recorder. Each open step has its fix.
Help on each page
The ? on top explains the page you are on. “New in this version” lists what changed.
Compatibility check
Each part of the app is On, Reduced or Paused, with the reason. An engine update pauses only what it breaks.
Engine update guide
A backup, the update in Terminal, the checks, the result. You start each step. The app never updates the engine by itself.
Jobs after an update
“5 of 5 jobs that ran passed their check”, or the names of the jobs that worked before and fail now.
A link that comes back
When the engine stops, the window stays open and says so. The link starts again by itself. A run that is silent for 10 minutes gets a notice.
Type the name to confirm
Run again, pause, archive, “never ask”: each needs the name typed first.
Updates
The app tells you when a new version is there. It installs only after you select it, and macOS checks that it comes from the maker of the app.
Backups
An edit of a profile file or a skill keeps the old version. Mesh Helm’s own data (rules, budgets, settings) gets a copy each day. Restore a copy, or export and import all of it as one file.
Off switch
One empty file stops the recorder and the guard.
Light and dark
System, light or dark.

In the app, you install skills and MCP servers from the catalog. The app has its own setup for the most used model providers and for Telegram, Discord and Slack. For the other kinds, on Hermes, it opens the setup of Hermes in a Terminal window; on Mesh, the Mesh app sets them up.

Two engines.

Mesh Helm runs your agents on Hermes Agent, or on Mesh, the engine inside Mesh Bot. Every feature on this page works on both, except the few named below. You pick the engine in Settings.

Hermes Agent

The free, open-source engine by Nous Research. Mesh Helm reads the Hermes folder you already have, and the checks run in a small plugin inside the engine. A Mac that used Mesh Helm with Hermes keeps Hermes.

Mesh

No Hermes needed. Each Mesh bot is a profile, and Mesh runs the checks itself: the Duplicate Guard, the Night Brake, the Watchdog and trial runs. A new install starts on Mesh when Mesh Bot is on the Mac.

What is different on Mesh: the Watchdog sends to a bot's Telegram chat; models and channels are set up in the Mesh app, not in a Terminal window; commands that Claude Code blocks are not carried over, because Mesh blocks whole tools; and the pages inside Hermes Desktop need Hermes.

Two ways to use it.

The checks run in one place: inside the engine. You look at the result where you like to work.

The Mesh Helm app

Its own window, with the chat in the middle and everything on this page around it.

It talks to the engine through the Agent Client Protocol, an open standard.

Some actions are in the app only: the budget stop, the job form, the MCP test, and profile copy, compare, export, import and delete.

The newest pages are in the app only too: Trial run, Review, Changes, MCP tool control, Context X-ray and the Night Brake settings.

Inside Hermes Desktop

With Hermes Agent: the same pages in the side list of Hermes Desktop: Home, Costs, Automations, Profiles, Map, Security, AI providers, From Claude Code, Health and alerts.

A strip in the title bar shows what runs, what needs you and the cost today.

Mesh Helm is not on sale yet. Leave your email address and you get one message when the Mac app is ready.

Get early access