NeufAgents · neufagents.com

Tools for the agent era.

Small, careful software for people who run AI agents — built in the open, shipped from day one.

New — open source

MCP Triage

mcp-triage v0.1.0

A triage for broken MCP setups. One command scans the config files of the eight agent clients on your machine, explains what is broken in plain English — and repairs the mechanical class (comments, trailing commas) with a parse-gated write and a backup.

  • Eight clients — Claude Desktop, Claude Code, Codex, Cursor, VS Code, Windsurf, OpenClaw, dsh
  • Checks that matter — broken JSON/TOML, commands not on PATH, missing env vars, relative paths, transport mismatches, cross-client drift
  • Conservative by design — only provably safe fixes are written; everything else gets an explanation and a hint, never a guess
GitHub ↗ npm ↗
MIT45 tests passingzero runtime deps
bash — triage 
# every MCP client on this machine
$ npx mcp-triage

MCP Triage v0.1.0 — scanned 3 config file(s)

Findings (4):
  [ERROR] Claude Desktop: trailing comma in JSON
           → remove the comma; restart the client
  [ERROR] Codex "notes-mcp": command not on PATH
  [WARN ] Codex "notes-mcp": relative path ./notes
  [INFO ] "filesystem" drifts across 2 clients

Summary: 2 errors, 1 warning, 1 info

# fix the mechanical class — with a backup
$ npx mcp-triage --fix
  [FIXED] Claude Desktop — removed trailing comma
Now shipping

dsh-healthcheck

dsh-healthcheck v0.1.1

A read-only health check for your DeepSeek Harness session library. It spots the two things that quietly kill long agent sessions — runaway size (OOM risk) and crash-torn logs (unclosed turns that resume silently repairs).

  • Library scan — totals, largest sessions, size & event-count warnings
  • Deep inspection — crash-tail detection, tool-call/result pairing, event distribution
  • Read-only — it never writes a byte of your data
GitHub ↗ Discussion ↗
MIT11 tests passingdsh 0.1.5-rc.2
bash — install 
# install the plugin
$ dsh plugin --profile web add \
    github:neufagents/dsh-healthcheck

# then ask your agent
> run a health check on my dsh sessions
Services

Deploy it. Fix it. Keep it running.

We work on self-hosted AI agent setups — Claude Code, OpenClaw, DeepSeek Harness, n8n and similar stacks, MCP configs included: if a client still refuses to start after a clean mcp-triage scan, send us the report — that is the kind of case we take. New servers deployed and hardened, broken ones repaired, healthy ones looked after. Fully async, everything in writing — and we take on a small number of projects at a time, so each one gets done properly.

Deploy & harden

Your agent stack, freshly set up on a clean server — and proven to boot.

  • Install on a clean Linux server, configured for real use
  • Hardening: non-root user, firewall, SSH keys only, automatic updates
  • Smoke tests and boot verification, shown in the handover
  • A short runbook: start/stop, logs, updates, recovery
from $149 · one-off
Fix a broken setup

Something stopped working — we find the root cause and fix it.

  • Installs that never finished, services that won't start, configs broken after an update
  • Remote diagnosis with a written plan; the fix applied and verified with logs
  • If it's outside what we can do well, we say so before anything is paid
from $99 · one-off
Managed ops

Monitoring, backups and updates for a stack that has to stay up.

  • Uptime, disk and service checks with alerts
  • Scheduled backups — tested and restorable, not just a cron job
  • Careful security updates, plus a plain-language monthly report
  • Month to month — no lock-in
from $89 / month
How it works
01

Write to us. A few lines: what you run (or want to run), where it lives, and what you need. Plain email — no forms, no calls.

02

Scope and price, in writing. We reply with exactly what we'll do, what it costs, and how access works — before anything is paid or touched.

03

Approve, then we work. You approve in writing; we do the work and hand over the proof — tests, logs, runbook.

Do you need access to my server?
Yes — SSH, ideally key-based, and you can revoke it at any time. We start read-only where possible, and on one-off jobs we close our access after handover.
Which stacks do you support?
Claude Code, OpenClaw, DeepSeek Harness, n8n and similar Node/Python agent setups. If yours isn't on the list, ask — we'll tell you honestly whether we're the right fit.
Can we do a call?
No — we work fully async by design. Everything in writing is faster to act on and keeps a clear record for both sides.
Do I need a server already?
No. Tell us your region and budget — we'll recommend a provider and specs. You pay the provider directly; we never resell servers.
How does payment work?
Once we agree the scope in writing, we send a secure payment link (card or bank transfer). Nothing is charged until you've approved the plan.

Tell us what you're running — we'll tell you how we can help.

[email protected]

Usually answered within a business day.