NEKKO LABS · OPEN SOURCE · MIT
NEKKOS
Train your models. Run your agents.
Own your machine.
An open-source, local-first AI coding & cowork app. Point it at Ollama, LM Studio, or vLLM in one click, or bring your own cloud keys, and let a model that can actually see and touch your codebases do the work.
No telemetry · no account · your code never leaves your machine unless you say so
SettingsView.tsx and wiring the theme store…01 · THE HEADLINE FEATURE
Local models, finally first-class
Most tools bolt local models on as an afterthought. Nekkos is built around them.
- Auto-discover a running Ollama, LM Studio, or vLLM on localhost
- Pull, load, and unload Ollama models from the app
- See what's resident in memory, and stop a server without leaving the app
- Manage servers remotely and watch token usage over time
- Bring cloud keys too, Anthropic, OpenAI, OpenRouter, any OpenAI-compatible endpoint
localhost:11434localhost:1234localhost:800002 · THE HARNESS
The harness is the other half of the model
Frontier labs ship raw capability. Almost everything else you feel as "this model is good in my codebase" is harness: what it gets shown, what it can reach, and what happens when it's wrong. Nekkos hands you all three, in one window that runs chat, cowork, and code.
What it sees
A short context window stops being the bottleneck once you decide exactly what goes in it.
- Context Inspector Every file, guideline, memory entry, and index snippet entering the next prompt, with per-source token counts, a window-headroom bar, and a toggle or pin on each. Your unsent draft is counted as you type it.
- Grounded retrieval A multi-folder index (file tree, code outline, search) so a large repo arrives as the relevant slice instead of a wall of text, across every folder you wire into the chat.
-
Guidelines and memory
AGENTS.md/CLAUDE.mdplus a markdown memory store, global and per project, with every entry individually switchable. -
A living spec
Build
SPEC.md(and a plan and task list) out of the conversation, then feed it back as context, so intent survives a window that can't hold the whole chat. - Prompt analyzer An offline, zero-latency read of your draft: parts detected, weak spots underlined, an A to F score, click-to-fill chips for whatever's missing, and a recommended model tier.
What it can reach
Capability the weights don't have to carry themselves.
- Sub-agents and swarms The agent delegates a scoped subtask to a fresh child session with its own clean context. Solo, balanced, or full swarm, with depth limits, and the whole tree visible live. The cheapest way past a small window is not filling it.
- Skills Procedures, not prompt snippets: research, plan, review, security-review, test, commit, PR, each rendered as a real workflow graph, plus a marketplace and one-click install.
- MCP, both directions Connect stdio or streamable-HTTP MCP servers and their tools merge into the agent loop, toggleable per chat. Nekkos is also an MCP server, so other agents can drive it.
- Connectors Linear, Slack, Discord, Gmail, and Drive, folded into context as snippets instead of pasted by hand.
- Goal loops that terminate Hand over a finish-this goal: it plans first, re-injects the plan every iteration, runs on a bounded window, and stops on a turn or time budget. A stall detector asks you for a hint instead of spiraling.
When it's wrong
Smaller models slip more often. The fix is making a slip cheap, not pretending it won't happen.
-
Guardrails
A risky-command classifier (destructive filesystem writes, force push,
curl | sh, privilege escalation) behind an allow / ask / deny policy you can edit. Ask, Guardrails, or YOLO, per chat. - Sandbox File access jailed to the folders you wired in, with optional Docker isolation for shell commands.
- Diff review Every agent edit lands in a Changes panel as a line diff. Keep or revert at line, file, or all, without leaving the window.
- Recovery, built in An empty reply mid-loop is retried with a nudge and then surfaced rather than silently swallowed, and "Reset here" rewinds the chat to any earlier message and runs again from there.
And while all of that runs, Nekkos drifts in zero-g, laps its own orbit, and bats at a passing comet.
03 · WHY NEKKOS
Not a chat box. Not a blind terminal.
LM Studio runs models. Nekkos runs with your work.
Local model UIs are a chat box around a model, no awareness of your files. Nekkos reads, edits, searches, and runs inside your real codebases: multi-folder index, inline editing, per-project memory, guardrails. Same one-click local setup, but the model actually does the work.
The power of an agentic CLI, with eyes.
Terminal agents are powerful but blind, you can't see what changed without git diff, and editing means leaving the tool. Nekkos gives an IDE-like surface: browse the tree, view files and diffs, edit inline, with every action visible through the Context Inspector and approvals.
04 · SAME APP, ANYWHERE
Run it your way
One engine and one UI across every runtime, pick what fits.
Desktop
Native app for macOS, Windows, Linux.
● AvailableSelf-hosted web
npx nekkos, the same UI in your browser, fully offline.
Docker
docker compose up, workspaces as volumes, local models via host network.
Nekkos Cloud Paid
Hosted, with an always-on Zero-Data-Retention mode, cloud history + files, and your phone driving your local model via a secure E2E relay.
○ PlannedNekkos Cloud keeps inference and tools on your machine, the relay is an end-to-end-encrypted pipe to your own paired agent, so using your local model stays private by design.
05 · TOUCHDOWN
Download Nekkos
Free and open source. We picked the build for your machine, change it if we guessed wrong.
All builds live on GitHub Releases.
Or build from source: git clone · npm install · npm run dev
macOS says it's "damaged"? It's just unsigned, run xattr -cr "/Applications/Nekkos.app" once, then open. Details