v1.2.0-rc|860 tests passing|5 platforms

Your server
shouldn't need
ChatGPT.

SSH client with a local AI copilot. Inspect logs, diagnose failures, preview commands, approve fixes. Everything runs on your device. Nothing leaves.

Download HAMMAView source →

No account required. No telemetry. No cloud dependency.

hamma — prod-api-3 — 10.0.1.42
SSH

$ hamma connect prod-api-3

Connected to 10.0.1.42:22 as deploy

› nginx is returning 502 on /api/health

Analysis

Upstream backend (port 3000) is not responding.

Process node (pid 4891) exited 4 min ago: OOM killed.

Proposed fix

risk: medium

sudo systemctl restart node-api

sudo journalctl -u node-api --since "5 min ago" -n 20

›approve|edit|reject

✓ node-api restarted. Health check passing.

›

The problem

Every time you paste a server error into a cloud AI, you leak production.

Internal IPs. Routing tables. Database connection strings. Auth tokens in stack traces. Your nginx.confgoes to OpenAI's training pipeline.

In compliance environments, that's a violation. In air-gapped networks, it's impossible. In most teams, it's just Tuesday — because there's no better option.

Meanwhile, SSH clients still assume you already know what to type. You don't need another terminal. You need a terminal that understands what went wrong.

What HAMMA does instead

→ AI runs on 127.0.0.1

→ Prompts stay on disk

→ Commands previewed before execution

→ Risk scored. You approve or reject.

→ Works offline. No account. No subscription.

Paste less. Leak less.

Approve every command.

Capabilities

Terminal when you need it. Buttons when you don't.

local ai

AI Copilot

Streaming local LLMs. Diagnoses failures from logs. Proposes commands with risk scores. You approve or reject.

# AI suggests:

sudo systemctl restart postgresql

⚠ risk:low — restarts db service

No "I'd be happy to help."
Root cause. Commands. Done.

ssh

Terminal

xterm-compatible. 256-color. Custom mobile keyboard row. Session reconnect. Scrollback restored.

sftp

Visual SFTP

Browse remote files. Edit in place. chmod, chown, sudo fallback. Syntax highlighting built in.

sys

Docker & Services

Container list. Live logs. Start/stop/restart. systemd units. Process viewer. All without typing docker ps.

vault

Encrypted Storage

AES-256-GCM at restArgon2id key derivationKeys decrypted to memory only during sessionBiometric unlock

multi

5 Platforms

LinuxmacOSWindowsAndroidiOS

Architecture

No cloud brain required.

AI traffic is hard-locked to 127.0.0.0/8 at the application layer. Not a setting. Not configurable. Your prompts physically cannot reach the internet.

Data flow

┌──────────────────────────────────┐
│  Your device                     │
│                                  │
│  Server error / log output       │
│         │                        │
│         ▼                        │
│  HAMMA app                       │
│         │ localhost:11434        │
│         ▼                        │
│  Local LLM (Ollama/llama.cpp)    │
│         │                        │
│         ▼                        │
│  Response → preview → approve    │
│                                  │
└──────────────────────────────────┘
    ▲
    │ Nothing exits this box.
    │ Ever.

The HAMMA model

Gemma 4 LoRA adapter. Trained on 1,500+ Linux sysadmin failure-solution pairs. Zero conversational filler. Outputs root cause + exact bash commands.

size

5.34 GB (Q4_K_M)

downloads

1,400+ in 48h

categories

39 failure types

hardware

Any dev laptop

$ quickstart

ollama run hf.co/xayrullonematov/hamma-gemma-4-devops-GGUF:Q4_K_M

HuggingFace →Also supports: Ollama, LM Studio, llama.cpp, Jan

Approval system

AI proposes. You approve.

The AI can suggest rm -rf /. HAMMA makes that terrifyingly obvious before anything runs.

proposed command

low risk

systemctl status nginx

Read-only. Inspects service state.

proposed command

medium

systemctl restart postgresql

Interrupts active connections. Brief downtime.

proposed command

critical

rm -rf /var/lib/postgresql/data

Destructive. Permanently deletes database files.

Vault

AES-256-GCM encryption. Argon2id KDF. Keys exist in memory only during active SSH session.

Network

AI hard-locked to loopback. No analytics SDK. No crash reporter. Zero bytes leave your device.

Execution

Every command requires explicit approval. Full audit trail. Risk scoring on rm, chmod, iptables, and 30+ patterns.

Workflow

Connect. Ask. Approve. Done.

1

hamma connect prod

Add server once. Connect with one tap or command.

2

› why is the API slow?

Describe the problem in plain English. Or paste a log.

3

[analysis: 3 commands]

AI reads logs, identifies root cause, proposes fix.

4

[approve / edit / reject]

Preview every command. See the risk score. Your call.

5

✓ fixed. audit logged.

Execution, output, and full audit trail stored locally.

Nothing runs without your confirmation. Nothing leaves your device.

Who it's for

Built for unstable SSH sessions,small VPSes, and real server work.

The 3AM server reboot

Your app is down. You're half-awake. You need the right command, not a man page. HAMMA gives you root cause and the fix.

The solo developer

2-5 servers. No DevOps team. You deploy, debug, and maintain everything. HAMMA is the second pair of eyes.

The school server

One Linux box. Student records. No IT department. The teacher who manages it needs buttons, not bash.

The hackathon VPS

48 hours. Fresh server. Nginx won't start. You need answers, not documentation rabbit holes.

The air-gapped facility

Hospital. Government. Research lab. No cloud API calls allowed. AI has to run locally or not at all.

The bad connection

Field deployment. Satellite internet. SSH drops every 5 minutes. HAMMA reconnects and restores your session.

Works where dashboards don't.

Comparison

How HAMMA compares

CapabilityHAMMATermiusSSH+ChatGPTiSH/Blink
Local AI (on-device LLMs)✓———
Fine-tuned DevOps model✓———
AI loopback enforcement✓———
Risk scoring before execution✓———
Works fully offline✓~~✓
5 platforms✓✓~~
Visual SFTP + sudo✓✓——
Docker & systemd panel✓~——
SubscriptionFreePaidFreePaid
Your prompts go toNowhereCloudCloudn/a

Status

Actively developed. Honest about what's done.

Currently in Beta Release Candidate. 860 tests passing. Shipping weekly.

Core Client

shipped

SSH, SFTP, terminal, vault

Local AI

shipped

Inference, HAMMA model, risk scoring

NL Ops

shipped

Command plans, previews, audit log

Built-in Engine

active

No Ollama dep — ships inside HAMMA

Modules

planned

Swappable specialist adapters

For the server you reboot at 3AM.

Local AI. Zero telemetry. Human-approved execution. Free.

Download HAMMAGitHubHAMMA ModelDiscord

No account. No credit card. No cloud dependency. Just download and connect.