16 FAQ

16 — FAQ

Updated: 2026-08-21

Frequently asked questions. Also see docs/FAQ.md for the original short FAQ shipped with the codebase.

🌟 General

What is Nokido ?

An Autonomous, Local-First AI Operating System with Neuro-Symbolic Governance. It orchestrates many specialized brains (LLMs, embedders, SNN, GOAP, AMI planner), owns its memory (persistent RAG), filters every cloud call (Semantic Firewall + Sovereign Membrane), and stays local-first by default.

Short version : a high-tech hospital where every brain has a specialty, every exchange is filtered, and the whole thing remembers everything 24/7 on your machine. See MANIFESTO.md for the full philosophy.

Why "Nokido" ?

It forges things — code, plans, decisions. Also a nod to Geordi La Forge (Star Trek TNG) — the chief engineer who makes the impossible work with local resources.

Who maintains it ?

user (@user). Solo project, open to contributors.

What's the license ?

AGPLv3-or-later. If you host Nokido as a service over the network, you must publish your modifications. Read CONTRIBUTING.md for the rationale.

Is Nokido stable ?

It's alpha. The alpha branch is active development, beta is release-prep, main will track stable releases. Pin to a tag for production-like usage.

🚀 Getting started

What hardware do I need ?

Minimum :

• 8 GB RAM.
• 4 CPU cores.
• 2 GB disk.

Recommended :

• 16+ GB RAM (32 if you want to run the ml extras with torch).
• 8+ cores.
• An integrated GPU (Radeon iGPU, Intel Arc, NVIDIA, M1+ Metal) for the brain_worker embedder.
• 10+ GB disk if you keep the RAG history.

Tested on an AMD Ryzen 7 8700G + Radeon 780M iGPU (consumer APU). All benchmarks in the README come from that setup.

Why Python 3.12 and not 3.11 ?

Several reasons :

• Modern type hints (X | None, dict[K, V]).
• tomllib in stdlib.
• Better async perf.
• The faiss-cpu wheels target 3.12.

Python 3.13 and 3.14 also work — pyproject.toml lists all three.

Do I need to know Python to use Nokido ?

No. The Docker setup gets you running in 5 minutes. You interact via your MCP client (Claude Code, Gemini CLI, Codex CLI, Cline) or via the TUI (nokido-cli).

To extend Nokido (new modules, custom skills), yes — Python is the primary language.

What's the difference between Nokido and Ollama ?

Ollama is a local LLM runtime — one process, one or more model files, an OpenAI-compatible API. Nokido uses Ollama as one of its 29 LLM providers.

Nokido adds :

• Multi-provider routing with cascade fallback.
• Persistent memory (RAG).
• Semantic firewall + sovereignty membrane.
• Multi-agent orchestration.
• Neuro-symbolic governance.
• AMI cognitive stack.
• A real RBAC.
• Sandboxing, vault, audit, etc.

Think of Ollama as the gas station and Nokido as the car.

What's the difference between Nokido and LangChain / AutoGPT / Agents SDK ?

The big inversion : in Nokido the hub is the orchestrator, the LLM is a transient worker. LangChain / AutoGPT make the LLM the orchestrator (which re-sends the full chain-of-thought each turn). Measured outcome : Nokido consumes 5-15× fewer tokens for the same end result. See MANIFESTO §3.5.

Nokido also has things they don't : Semantic Firewall, Sovereign Membrane, AMI cognitive stack, 6-ring RBAC, neuro-symbolic verdict path. The closest comparison is OpenHands (formerly OpenDevin), but even there Nokido differs : sovereignty + memory + governance.

🤖 LLMs & providers

Which LLM should I use ?

For most tasks, auto routing : the cascade picks for you (local first, free cloud next, paid only if you've enabled it). To force :

• Code : qwen2.5-coder:32b via Ollama (local), or groq (free cloud, fast).
• Reasoning : gemini_pro (free tier 1k/month), or cohere_command_r_plus.
• Vision : gemini_flash (1M context, free).
• Long context : gemini_flash again, or cohere_command_r_plus (128k).
• Tool calling : llamacpp_local (qwen2.5-coder + --jinja), or ollama.

Do I have to use any cloud LLM ?

No. The local tier (Ollama, llama.cpp, LM Studio, brain_worker) covers most workflows. Cloud is a quality boost when local isn't enough.

What about privacy when calling cloud LLMs ?

Two layers automatically applied :

1. SemanticFirewall.pre_flight : DLP + injection check + canary.
2. SovereignMembrane.wrap : HMAC-aliased anonymization of hostnames, paths, IPs, tokens, UUIDs.

The cloud provider sees [host:xxx] and [path:yyy], not your real data. The hub restores aliases in the response transparently.

You can also disable cloud entirely : delete the API keys from the vault, and the cascade will only use local providers.

How do I add a custom provider ?

See 05 — LLM providers § Adding a custom provider.

Why is my Groq / Cerebras / Cohere call failing ?

The most likely cause is quota. Check :

http://127.0.0.1:8766/admin/providers

If the percentage is red, the provider is skipped by the cascade until the quota resets (start of next month/day).

🛡️ Security

How are secrets stored ?

In a machine-wide OS-encrypted vault :

• Windows : DPAPI CRYPTPROTECT_LOCAL_MACHINE.
• macOS : Keychain.
• Linux : libsecret / Secret Service.

See 08 — Vault & secrets.

What if the vault is leaked ?

The vault file (data/machine_vault.dat) is encrypted with the machine key. A network attacker who pulls the file cannot decrypt it without local access.

If they have local access (e.g. compromised user account), they have the same threat profile as a .env on disk. The vault helps against passive leak (backups, git push, file sync), not against active compromise.

Can I expose the hub to the internet ?

Don't. The hub binds 127.0.0.1 by default. To access from another machine, use Tailscale or WireGuard. Do not change the bind to 0.0.0.0.

How do I report a vulnerability ?

Private disclosure : Security tab → Report a vulnerability. See SECURITY.md. We aim for 72-hour acknowledgement.

Are there backdoors ?

No. The code is AGPLv3 — fully readable, fully auditable. The hub binds localhost, has no telemetry, no "phone home" pings. Pre-commit hooks (Gitleaks + forge_secret_guard) prevent accidental secret inclusion.

If you want extra paranoia : build locally, audit app/forge_*.py, inspect network_log for outbound calls.

💾 RAG & memory

How big is the RAG ?

Default fresh install : ~3k chunks (seed data — lessons, configs, ADRs). After a few weeks of active use : 50-100k. The reference dev setup is ~531k chunks (warm tier).

Can I delete the RAG and start over ?

Yes :

rm RAG/embeddings.db
python tools/forge_db_bootstrap.py  # re-import baseline

You lose all learned lessons and indexed content.

How do I query the RAG ?

Three ways :

• MCP : rag tool with action=search, topic=<query>.
• Web UI : http://127.0.0.1:8766/forge/rag.
• SQL (ring 0 only) : query tool with FTS5 syntax.

Can I export the RAG ?

Yes. tools/forge_db_seed_export.py dumps tables to JSONL with schema. Useful for backups and cross-machine sync.

🔌 MCP & clients

What is MCP ?

The Model Context Protocol — Anthropic's open standard (2024) for client-LLM tool interaction. Nokido implements MCP 2025-03-26.

Which clients work with Nokido ?

Tested : Claude Desktop, Claude Code, Gemini CLI, Codex CLI, Cline (VS Code), MCP Inspector, custom HTTP clients. Any MCP-compatible client should work — see 04 — MCP clients setup.

Can I write my own MCP client ?

Yes. Speak JSON-RPC 2.0 over HTTP or STDIO with the schema from 06 — Hub API reference. 25 tools available.

⚙️ Operations

How do I update Nokido ?

cd Nokido
git pull
# Docker
docker compose -f docker/nokido/docker-compose.yml build
# Native
EXTRAS=$YOUR_PROFILE bash install.sh

How do I back up everything ?

Critical paths to back up :

• RAG/embeddings.db — the memory.
• data/machine_vault.dat — vault (Windows).
• Nokido.env — config (no secrets).
• seed/*.jsonl — bootstrap data.
• logs/lessons_learned.md — human-readable lessons trace.

tar czf laforge-backup-$(date +%F).tar.gz \
    RAG/embeddings.db data/machine_vault.dat \
    Nokido.env seed/ logs/lessons_learned.md

How much does Nokido cost to run ?

• Local-only : $0 / month.
• With free-tier clouds : $0 / month if you stay under quotas.
• With paid APIs (Anthropic / OpenAI direct, etc.) : depends on use. Typical dev session : $0.50–$5 / day if you lean on Claude Sonnet for hard tasks.

The cascade is designed to keep cost minimal — local first, free next.

Can I run Nokido headless ?

Yes. Don't launch the TUI ; just run laforge-hub (or the Docker compose). All functionality is available via HTTP MCP.

Can I run multiple Nokido instances on the same machine ?

Yes — change the port mapping. The hub takes LAFORGE_HUB_PORT env var. Each instance needs its own RAG/, data/, logs/, sandbox/ directories.

🧪 Development

How do I add a new tool to the hub ?

1. Register it in app/forge_mcp_registry.py::ToolRegistry.
2. Add the dispatch handler.
3. Add it to tools/hub_middleware.py::_ALLOWED_TOOLS.
4. Test : pytest tests/test_forge_mcp_registry_namespace.py tests/test_forge_mcp_registry_unicode.py.

How do I write a custom skill ?

See docs/skills/nokido/SKILL.md. Skills are markdown files describing capabilities. The skill MCP tool loads them into the RAG and exposes them to agents.

How do I run tests ?

pytest tests/                   # all
pytest tests/ -m unit           # fast tests only
pytest tests/ -m security       # security suite
pytest tests/test_forge_scorecard.py   # one file

CI matrix : Ubuntu + macOS + Windows. See .github/workflows/ci.yml.

Where do I file an issue ?

GitHub : https://github.com/Nokido-labs/nokido/issues. Use the Bug report or Feature request template. For security : private email (see above).

🛣️ Future

What's the roadmap ?

See MANIFESTO §9 and 13 — Hardware roadmap. Summary :

• Phase A (current → 12 months) : NPU edge accelerator support.
• Phase B (12-24 months) : neuromorphic backend (Akida, Loihi).
• Phase C (24-48 months) : analog CIM / RRAM / photonic.
• Phase D (48+ months) : continuous neural grids.

Will there be a SaaS version ?

Not in the foreseeable future. The whole point of Nokido is local sovereignty. A SaaS Nokido would be self-contradictory.

Will there be Windows ARM / Apple Silicon support ?

Apple Silicon : already works (M1/M2/M3) — CoreML backend for brain_worker, Metal for torch. Windows ARM : not tested ; PRs welcome.

Can I use Nokido with my Mac without Docker ?

Yes : bash install.sh on macOS works natively. See 01 — Installation § macOS.

Will you support more LLM providers ?

Yes — but vendor-neutrally, via litellm. If litellm supports it, Nokido does (add the spec in forge_provider_specs.py). If a provider needs a custom client (Anthropic SDK, mistral_common…), file an issue.

🤝 Community

Where do I get help ?

• This wiki.
• GitHub Discussions (once enabled on the public repo).
• Mention @user for design questions.

Can I sponsor / donate ?

Public release will set up GitHub Sponsors. Until then, contributing PRs is the best way to support the project.

How do I cite Nokido in academic work ?

@software{nokido_2026,
  author = {user},
  title  = {Nokido: An Autonomous, Local-First AI Operating System
            with Neuro-Symbolic Governance},
  year   = 2026,
  url    = {https://github.com/Nokido-labs/nokido}
}

Replace the URL with the canonical citation URL once a release is tagged.