Helmies
AI agents

Not one agent. A team of them.

Helmies-agent runs operations + marketing. Hermes takes phone calls + voice. OpenClaw ships code from your repos. All three share the same memory and tool registry, RLS-scoped to your tenant. Bring your own via MCP.

helmies-agent · supervised

Sonnet 4.6 · 142 tokens · €0.011

supervised
W
Wael · just now
Plan next 2 weeks of social posts. Spring colours, no buzzwords.
helmies-agent
brand.read…
trends.scan…
content.draft…
social.queue…
The roster

One memory. Three specialists.

Each agent has its own personality and tool set, but they share long-term memory and your tenant. Tasks can hand-off between them — Helmies-agent for ops, Hermes for the call, OpenClaw for the code.

Helmies-agent
Operations · marketing · finance

The default. Reads every row of your business. Picks Claude / GPT / Gemini per task. Three autonomy modes.

customers.listinvoice.draftmeta_ads.pausecontent.draftemail.send
Hermes
Voice · phone calls · autonomous

Phone-first. Answers calls, books meetings, summarises conversations into the CRM. Runs scheduled prompts via cron.

call.answertranscribesummariseschedule.queuesms.send
OpenClaw
Code · repos · pull requests

Connect a GitHub repo. OpenClaw reads, plans, writes the code, runs your tests, opens a PR against your branch.

repo.readbranch.createcode.writetests.runpr.open
MCP
Bring-your-own agents

Any MCP server you trust. Tools auto-exposed to the registry. Tenant-scoped permissions.

mcp.notionmcp.figmamcp.linearmcp.githubmcp.your-server
Try it

Pick a scenario. Watch it execute.

A static demo of the agent surface. Inside Helmies every call is logged, every action is reversible, and every tool can be scoped per role.

Autonomy
Scenarios
helmies-agentmode: supervised connected to 18 tools
W
Wael
Pause Meta campaigns where CTR < 0.4% and reallocate budget to the top 2.
helmies-agent · thinking
01meta_ads.listok
02meta_ads.pauseok
03meta_ads.budget_setok
04alerts.createok
summary

Paused 3 campaigns burning €420/day at 0.21–0.38% CTR. Moved €240/day into Northern Lights Pro and €180/day into Spring Bundle (top performers, CTR 1.6%+). Set a guardrail to alert if combined ROAS dips below 2.1.

⌘↵
Multi-model router

Right model. Right job.

We benchmark prompt → quality → cost continuously and switch the agent to the best model per tool. You see the margin per model in your usage dashboard.

claude · 1M ctx$$
Opus 4.7
Hard reasoning, code, ambiguous prompts
claude$
Sonnet 4.6
Default for writing, summaries, sequences
claude¢
Haiku 4.5
Batch ops, classification, cheap+fast
openai$$
GPT-5.5
Long-context, RAG over big docs
google$
Gemini 2.5
Multimodal, image understanding
groq¢
Llama 3.3 70B
Ultra-fast cheap fallback
fal.ai¢
flux-dev
Image generation for social
replicate¢
SDXL
Image generation, ControlNet

Model router

Auto-pick by tool · token meter visible

anthropic
Opus 4.7 (1M)
Hard reasoning, code, long context
$$
28%
on
anthropic
Sonnet 4.6
Default writer + sequence drafts
$
55%
on
anthropic
Haiku 4.5
Batch classify, triage, low-stakes
¢
13%
on
openai
GPT-5.5
Backup for RAG over big docs
$$
5%
on
google
Gemini 2.5 Pro
Multimodal · image understanding
$
4%
on
fal.ai
flux-dev
Image generation for social
¢
3%
on
groq
Llama 3.3 70B
Ultra-fast cheap fallback
¢
1%
off
Tool registry

The agent has 80+ tools and counting.

Auto-derived from the code. Every tool is documented in /admin/api-docs the moment it ships. Tools are scoped per role and rate-limited per tenant.

Customers
  • customers.list
  • customers.create
  • customers.update
  • customers.merge
  • customers.history
Projects
  • projects.list
  • projects.create
  • projects.tasks
  • projects.milestones
  • projects.profitability
Money
  • invoice.draft
  • invoice.send
  • invoice.list
  • quote.create
  • quote.accept
  • payroll.run
  • taxes.file
Marketing
  • leads.search
  • leads.enrich
  • sequence.create
  • ads.list
  • ads.pause
  • content.draft
  • social.queue
  • trends.scan
Comms
  • email.compose
  • email.send
  • email.list
  • slack.notify
  • calendar.add
  • calendar.find_slot
Files & data
  • files.upload
  • files.search
  • files.share
  • sql.read
  • sql.export
Code (OpenClaw)
  • repo.read
  • branch.create
  • code.write
  • tests.run
  • pr.open
  • pr.review
Voice (Hermes)
  • call.answer
  • call.dial
  • voice.transcribe
  • voice.summarise
  • sms.send
Hand-off

When the agent needs a shell. It opens one.

For code-ish tasks, the agent can drop into the in-app terminal, SSH to your server, and run Claude Code there. You see the session live, can stop it any time, and it's logged.

  • Read-only by default
    Write requires explicit consent.
  • Sandboxed
    Per-tenant SSH key vault, no cross-tenant access.
  • Audited
    Every command + diff stored, 90 days.
  • Recoverable
    Stop at any time; revert via git or backups.

Terminal · ssh helmies@deploy

ttyd + tmux · pane 1 · Claude Code

connected
›

Stop tab-switching. Start delegating.