01 · Project memory
It remembers — and shows its work.
Every fact Nemocode keeps has proof behind it and a trust score. Nothing is stored until you say yes.
It remembers
everything.
Nemocode writes down what it learns about your codebase, with proof and a trust score for every fact, and brings it back the moment it opens the file.
Every other agent
starts over every single morning.
It re-reads your repo. It re-guesses your conventions. It makes the same mistake you corrected on Monday — and you pay for the rediscovery, in tokens and in patience.
02 · MCP
Your other agents get the memory too.
nemocode integrate claude|codex|opencode|claude-desktop gives them 15 tools to read, search and propose. Their proposals wait for your yes as well.
03 · Team sync
One memory for the whole team.
A Source shares the project memory over HTTPS with a pinned certificate. Members read, propose or review — and an audit log records every step.
04 · Skills
It writes its own skills.
After a long mission the curator turns what worked into a SKILL.md — as a proposal you approve.
05 · /goal
Set a goal. It keeps going.
/goal <condition> — Nemocode works turn after turn until a checker says it is met, or honestly says it cannot be.
06 · Control
Power, with a leash.
Five modes from read-only to auto. A sandbox enforced by the operating system, denied and read-only folders, and /undo for every change.
07 · Toolbox
The full toolbox.
Subagents, language servers, web search, hooks, a git worktree per session and scheduled tasks.
Everything in one agent.
The graph
A picture of what
your project actually knows.
This is the shape Nemocode builds for a codebase: the code indexed underneath, and above it the handful of things worth actually remembering — each with a confidence you can argue with.
33
kinds of evidence
3
trust levels: fact, likely, hypothesis
0
memories stored without your yes
1
command to start
One mission
What it looks like
in your terminal
> a refund posted twice last night — fix it Looked up (search: refunds, nightly job) └ 0.62 The nightly job is not idempotent — run twice, it double-counts refunds. └ 0.91 Money is stored in integer cents. Read 1 file (ctrl+o to expand) └ billing/nightly.py■ Update(billing/nightly.py) └ 88 - rows = pending_payouts(cur) 88 + rows = claim_payouts(cur, skip_locked=True) … +11 lines (ctrl+o to expand)■ Tests(pytest tests/billing -q) └ 41 passed in 6.2sFixed. The nightly job now claims each payout row before it pays it out, so a second run skips rows that are already taken.■ Untangled spaghetti for 1m 48s■ √ Harvest done — 2 proposals, 1 waiting for your click (/pending)> ►► accept edits on (shift+tab to cycle) · / commands · ↑↓ history · shift+↑↓ scroll · ctrl+o expand
Yours
On your machine.
Under your control.
Server, terminal and graph run locally. Your code is edited where it lives. The only thing leaving is the prompt your own provider key pays for.
Your code never touches our servers
It goes only to the model provider you pick — or nowhere at all, with a local model.
Every change is reversible
/undo and /redo step through every file change. Your git history stays yours: it commits only when you ask, and never pushes.
You hold the gate
It proposes what it learned; it never writes its own memory. Rules and decisions wait for your yes.
Anthropic · OpenAI · OpenRouter · Ollama · DeepSeek · Groq · Cerebras · and any OpenAI-compatible endpoint
Compared
It learns. The others take notes.
Claude Code, Codex and opencode are good tools — and they all start each session close to zero. Nemocode keeps a project memory you can check. Each dot is taken from the tool’s own documentation, as of October 2026.
| Nemocode | Claude Code | Codex CLI | opencode | |
|---|---|---|---|---|
| Memory and learning | ||||
| Learns from your work on its own | ● | ● | ● | ○ |
| Turns a solved hard problem into a skill, on its own | ● | ○ | ○ | ○ |
| Every fact has proof and a trust score | ● | ○ | ○ | ○ |
| Nothing is stored before you approve it | ● | ○ | ○ | ○ |
| Facts are linked in a graph | ● | ○ | ○ | ○ |
| Notices when the code behind a fact changes | ● | ○ | ○ | ○ |
| Opens a file and already knows what matters there | ● | ● | ○ | ○ |
| Projects can share their memory, if both agree | ● | ○ | ○ | ○ |
| Browse and edit the memory in a graph view | ● | ○ | ○ | ○ |
| Other agents can use its memory (MCP) | ● | ○ | ○ | ○ |
| Everyday tools | ||||
| Plan mode: read first, edit after approval | ● | ● | ● | ● |
| Undo or rewind file changes | ● | ● | ○ | ● |
| Subagents | ● | ● | ● | ● |
| Skills | ● | ● | ● | ● |
| Hooks | ● | ● | ● | ● |
| MCP servers | ● | ● | ● | ● |
| OS-level sandbox for the shell | ● | ● | ● | ○ |
| Language servers (LSP) | ● | ● | ○ | ● |
| Web search and fetch tools | ● | ● | ● | ● |
| A separate git worktree per session | ● | ● | ● | ○ |
| Scheduled tasks | ● | ● | ● | ● |
| IDE extension | ○ | ● | ● | ● |
Buy it once.
Keep it forever.
Solo
Price on request · one time
One developer, one machine at a time.
- The full agent and the full graph
- Every provider, your own keys
- 12 months of updates
Pro
most popularPrice on request · one time
For people who live in the terminal all day.
- Everything in Solo
- Unlimited projects
- 12 months of updates
Team
Price on request · per seat
One licence, several seats, the same conventions enforced.
- Everything in Pro
- Up to 10 devices on one licence
- Seat management and device binding
Straight answers
The questions you'd ask anyway
Does my code leave my machine?
The parts a task needs do — not to us, but to the AI provider you choose and pay for. Nemocode never routes them through our servers and we never see them. If you want nothing to leave at all, point it at a local model and it works the same way. The full detail is in the privacy policy, section 2.
Which models does it work with?
Anthropic, OpenAI, OpenRouter, Ollama (cloud or local), DeepSeek, Groq, Cerebras, and any OpenAI-compatible endpoint. You can even give each role its own model — a strong one for the main agent, a cheap one for the mechanical work. Model costs are yours and are not part of the licence.
What happens when the update window ends?
Nothing stops. The licence is perpetual and its token carries no expiry — a perpetual grant that dies when the network does is not perpetual. You keep running the last build you were entitled to, forever. Newer builds need an upgrade.
Can I get my data out?
Yes, and that is deliberate: the export works even with an expired or blocked licence, because your project knowledge is yours. It all lives in ~/.nemocode/ on your own disk anyway.
Windows?
Yes, natively on Windows 10 and 11 (x64). One command in PowerShell: irm https://nemonix.ai/nemocode/get.ps1 | iex. The same command updates it later.
Private beta is open
Small cohorts, real repositories. Tell us what you work on and we'll send an invite code and a founder price.
Request access[email protected] · Support: [email protected]