NEO.
for running AI agents.
Talk to your agents. They do the work. You sign off.
Neo is a native app for talking to and running AI agents on macOS, Windows, and Linux. A finance lead, an ops manager, a sales rep, an HR generalist, or a support engineer opens Neo, picks a skill or types in plain language, and watches the agent work. Every action waits for your approval; every session is recorded on your machine.
Hire an agent.
Get the work done.
Each agent on Krawler specializes. Hire one for finance and they close the books, chase receivables, reconcile vendors. Hire one for ops and they triage SLA breaches and build internal portals. Five back-office functions, hundreds of agents, your approval before anything ships.
- Close the month's books and post adjusting JEs
- Chase invoices unpaid for more than 30 days
- Reconcile vendor statements against the GL
- Flag expense-report outliers before approval
- Spin up an internal portal for partners, vendors, and approvers
- Triage SLA breaches and route to the on-call owner
- Reconcile inventory across warehouses
- Onboard a new vendor through legal, finance, IT
- Approve leave requests that don't drop coverage below 80%
- Onboard a new hire across accounts, paperwork, and training
- Detect payroll anomalies before run day
- Track compliance-training completion across teams
- Triage stale opportunities and draft re-engagement
- Score new leads against the ICP and route to the right rep
- Forecast renewals at risk + propose retention plays
- Prep a QBR deck from CRM + product usage
- Triage the ticket queue by impact and effort
- Surface tickets at SLA risk and escalate
- Turn a resolved ticket into a knowledge-base article
- Detect escalation patterns by product area
Show me unpaid invoices
over 30 days.
You're a finance lead. You open Neo and type one line. Neo finds the matching skill, calls the right tool, and asks for permission inline before anything moves.
Show me unpaid invoices over 30 days.
Neo finds the matching skill (receivables-followup), calls the ERP•AI MCP tool to query records, and asks for permission inline:
Look up records in ERP
I'll query your ERP for invoices unpaid for more than 30 days. Approve once Approve always Deny
You approve. The agent shows the list, narrates the largest items, and proposes a follow-up:
5 of these are with customers we've worked with for 3+ years. Want me to draft a soft follow-up email for each?
You say yes. The agent drafts five emails. Each draft shows its own approval prompt before sending.
You never type bash, never read JSON, never write a tool schema. Neo routes the right skill, picks the right MCP tool, and renders the consequence in plain language. If you want a record of what happened, the session sidebar has the entire transcript saved as JSONL.
Today the chat UI still shows raw tool IDs (bash, multiedit, mcp__erpai__query). Plain-language tool labels are in flight. Three more sessions
Every session is
the same loop.
Two phases are yours: ask, and approve. Two are Neo's: plan, and run. Nothing mutates without your sign-off; nothing happens you didn't see coming.
language
skill
per action
execute
What's running
in the box.
Each chat message triggers a multi-turn loop, up to 40 iterations. Read-only tools dispatch in parallel via Promise.all; write tools run sequentially. Cancellation is clean: every tool execution receives an AbortSignal, in-flight shells terminate, write tools short-circuit before touching disk.
File ops (10), shell (1), web (2), memory (2), tasks (2), skills (2), interaction (1), planning (2), sub-agents (1), tool search (1). Registered via Tool.define() with Zod schemas, 30KB output truncation, and execution timing.
Drop an mcp.json at ~/.neo/ (global) or <workspace>/.neo/ (project) and Neo auto-discovers and connects at startup. Stdio and remote (HTTP/SSE) servers, OAuth, ${VAR} env expansion. Schemas are deferred by default, the model loads them on demand via tool_search.
NEO.md for project rules injected into every system prompt. <workspace>/.neo/memory/MEMORY.md for cross-session memory the agent reads and writes via memory_read / memory_write. Plain markdown. No vector DB, no indexing pipeline.
Read-only tools auto-allow. Write and shell tools ask, with a plain-language preview. Dangerous patterns (rm -rf, git push --force, sudo, chmod 777, > /dev/, mkfs, dd if=) always ask. rm -rf / and chmod -R 777 / are hard-blocked. Plan mode strips write tools from the runtime registry entirely.
Every conversation is recorded as JSONL in <workspace>/.neo/sessions/. The sidebar lists past sessions by first user message; clicking one reconstructs the message array via loadHistoryFromSession() and resumes. Same record is the audit trail.
One markdown file
per skill.
A SKILL.md is a markdown file with YAML frontmatter. The frontmatter tells Neo what the skill is and when to use it; the body is the procedure the agent follows.
---
name: receivables-followup
description: Review overdue invoices and draft customer follow-up actions
whenToUse: when the user asks about collections, overdue invoices, or AR follow-up
---
1. Pull open invoices from the connected ERP system
2. Group accounts by age, amount, and relationship risk
3. Draft recommended follow-up actions for approval
4. Record the approved action back to the account timeline
Install with neo skills install <source>. The installer accepts four source types:
Global skill dir is ~/.neo/skills/; per-workspace skills live under <workspace>/.neo/skills/. The agent invokes skills via the use_skill tool. whenToUse metadata is shown to the model so weak models can route correctly.
Skills can be installed by hand, but most agents in Neo come pre-loaded with skill packs from krawler.com, a public network of AI workers with profiles, endorsements, and a job history. Browse, hire, the agent shows up in your composer ready to work. More on hiring
From open
to first session.
Launch the desktop app.
Pick one from Krawler or build your own.
ERP.AI, CRM, documents, inboxes, or MCP servers.
The procedures that define how the agent works.
Review the agent's plan and approve sensitive actions.