NurCLI MCP server
Commit Nov 11, 2026A cost-efficient coding agent CLI that uses a custom Rust harness to manage multi-provider LLM sessions and token usage.
Also known as nuroctane/nur-cli
- Stars
- 5
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 11, 2026
- License
- GPL-3.0
What is NurCLI MCP?
NurCLI is a cost-efficient coding agent CLI that uses a custom Rust harness to manage multi-provider LLM sessions and token usage.
Summary generated from the project's README
About NurCLI MCP server
Managing coding tasks with high token efficiency is what developers achieve using this Rust-based CLI agent. It functions as a personal coding assistant that optimizes costs by preventing duplicate reads, managing tool result spills, and preserving provider prompt caches. The system supports over 72 providers, including OpenAI, Anthropic, Gemini, and local engines like Ollama or LM Studio, allowing users to switch between models or use economy subagents that run on the cheapest available tool-capable model.
Users interact with the agent through a dense terminal user interface that features live sidegraphs, inline diffs, and native vision capabilities. Setup is handled via a single command using npx or platform-specific installers for Windows, macOS, and Linux. The tool manages credentials through a unified vault and supports various authentication methods, including browser-based sign-in for services like Cursor and GitHub Models. It also includes specialized tools for semantic code search, structural search and rewrite, and automated context compaction using TypeSafe judgments.
What you can do with it
- 1Optimize LLM token spendingA developer can reduce costs by using the harness to supersede duplicate reads and compact context using OMP-style response reserves.
- 2Execute multi-provider coding tasksYou can instruct the agent to use a specific model or switch providers to utilize subagents running on cheaper models.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablenuroctane/nur-cli is public and reachablegithub.com/nuroctane/nur-cli ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)GPL-3.0spdx.org/licenses/GPL-3.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-11github.com/nuroctane/nur-cli ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubNurCLI
Extremely efficient token spend
Use fewer paid tokens by default.
The harness supersedes stale duplicate reads, spills large tool results, preserves provider prompt caches, and compacts repeatedly using an OMP-style response reserve instead of a premature fixed percentage. It also supports local context estimates when providers omit usage, prompt caching, local models, and economy subagents that run on the cheapest tool-capable model your account has.
Tool-result previews include their spill path in the size budget, including long paths and tiny limits.
TypeSafe · Jev judgments run inside the harness: batched decisions help assess tool calls and results and prune context. Models can request compaction using typesafe with action = "compact"; automatic context recovery also tries the configured decision engine first and summarizes if needed. Provider-agnostic, with supported local engines including openJev-verdict-2.0, Bespoke-Nimble-9B, and Laya Core ML (docs/jev-local.md). Native failover and cross-provider subagents report their actual token and cost impact in Nur's session budgets and /usage totals.
Your personal coding agent. Custom Rust harness, dense gold TUI, native vision, tools, knowledge stack, gated execution. Multi-provider /provider plus a unified /auth vault. Any model via --model / /model / config.
nur # interactive gold TUI
Install: dead simple
One command on every OS:
npx nur-cli
Downloads the prebuilt native binary from GitHub Releases (no Rust toolchain, no clone, no build), drops it in ~/.local/bin, and runs the full stack setup. Keep it permanently with npm i -g nur-cli.
Prefer a one-liner without Node? The classic one-liners (build from source) or the Windows EXE (prebuilt) each drop nur on your PATH, pull in every runtime dependency they can, and wire the full agent stack before the TUI opens.
irm https://raw.githubusercontent.com/nuroctane/nur-cli/main/install.ps1 | iex
curl -fsSL https://raw.githubusercontent.com/nuroctane/nur-cli/main/install.sh | bash
After install
nur auth login # key -> ~/.nur/auth.json (or set NUR_API_KEY)
nur # open the TUI
nur doctor # health check
Or run nur and use /provider in the TUI (/login remains an alias): pick any of 72 providers
(OpenAI, Anthropic, Gemini, xAI, Groq, OpenRouter, Refuseless, local Ollama/LM Studio, Meta Model API, and so on).
For Refuseless, set REFUSELESS_API_KEY or run nur auth login --provider refuseless; /model lists its available models.
For OpenAI, Anthropic, xAI, Kimi Code, Google Gemini, Azure OpenAI, GitHub Models, GitHub Copilot, Cursor, Factory Droid, Cline, Muse Code, DeepSeek Harness, and ZCode (GLM) you can sign in with a browser or an official CLI session (or import an existing Codex / Claude Code / Grok / Kimi / Cursor Agent / Droid / Cline / Muse / dsh / ZCode session) instead of pasting an API key. nur auth login --provider muse --browser (and the same for deepseek / zcode) is the CLI equivalent of TUI /login. OAuth sessions refresh before use, retry once after an authentication rejection, and /model detects the models available to the active credential. Cursor uses cursor-agent login and runs chat through the Agent CLI (no API key required); Factory Droid works the same way through droid exec (NUR_DROID_BIN when the launcher is outside PATH). DeepSeek's official harness is API-key only ($DSH_HOME/.credentials.yaml). No credential on launch and login opens automatically.
Long turns are safe on every provider: no clock cuts off a response that is still streaming, and a dropped connection (reset, stale keep-alive, edge restart, or 300s of silence, NUR_PROVIDER_IDLE_TIMEOUT_SECS) reconnects and replays the request with backoff, retracting any partial text first. Esc cancels immediately; NUR_PROVIDER_TURN_TIMEOUT_SECS sets an optional hard cap. On Windows, Nur also recovers successful Cursor Agent turns from Cursor's durable transcript when affected CLI builds emit no redirected stdout.
Keys are kept per provider, so switching with /provider never strands the one you switched away from - it stays available for failover, fusion, and subagents running on another provider's model. /auth opens one scrollable vault for every provider, where you can add, replace, import CLI/OMP, or delete credentials without switching the current chat. Imported OMP credentials retain their real kind, so API keys use API endpoints and OAuth subscriptions use their provider-specific OAuth routes. See docs/authentication.md.
Sign-in handling and its regression coverage are documented in the 65-provider authentication review, including vendor CLI progress, token rotation, credential replacement, and Claude's manual fallback.
Update
nur update
Pulls latest main when a Laboratory checkout exists (~/laboratory/nur-cli or ~/Laboratory/nur-cli), rebuilds, reinstalls nur on PATH, re-provisions the ecosystem. No checkout? Falls back to nur install.
| Also fine | |
|---|---|
Re-run npx nur-cli |
Fresh prebuilt binary + stack |
| Re-run the one-liner | Full rebuild from GitHub |
| Re-download + double-click Windows EXE | Prebuilt path |
nur install |
Reinstall this binary + stack (no git pull) |
Verify: nur --version · nur doctor. Detail: docs/setup.md → Update.
Other ways to install
- Open Releases → latest
- Download
nur-windows-x86_64.exe - Double-click it (or
.\nur-windows-x86_64.exein a terminal)
Copies itself to ~\.local\bin\nur.exe, PATH, prereqs, ecosystem + browser setup, then opens NurCLI.
3. Already cloned
cd nur-cli
.\install.ps1 # Windows
# ./install.sh # macOS / Linux
4. Manual cargo build
git clone --depth 1 https://github.com/nuroctane/nur-cli.git
cd nur-cli
cargo build --release
./target/release/nur install # Windows: .\target\release\nur.exe install
nur auth login
What install puts on your PC
Everything is local. Secrets never go into the git repo.
A. Tooling install may add (if missing)
| Piece | Why |
|---|---|
| Rust / cargo | Builds the CLI (one-liner / cargo only; npx path needs none) |
| Git | Clone / update |
| Node.js 20+ | PLUR · Ruflo · Executor · skills · browser · AKM (the npx shim itself needs only Node 18+) |
| Bun | AKM runtime |
| uv | Graphify |
ripgrep (rg) |
Fast grep / glob, and the file walk behind sgrep |
| sgrep | Semantic code search (sgrep tool); built with cargo install --git https://github.com/context-dot-dev/sgrep --locked, model (32 MB) downloaded once |
| ast-grep | Structural search and rewrite (ast_grep tool); npm i -g @ast-grep/cli (prebuilt), or cargo install ast-grep --locked |
| ffmpeg | extract_frames |
B. NurCLI itself
| Piece | Path |
|---|---|
nur binary |
~/.local/bin/nur (Windows: nur.exe) |
| SHA-256 record | ~/.local/bin/nur.sha256 |
| Source checkout (one-liner) | ~/laboratory/nur-cli (override NUR_CLI_DIR) |
| User PATH | ~/.local/bin |
C. Data home: ~/.nur/
| Path | Purpose |
|---|---|
auth.json |
Active credential |
provider_keys.json · provider_sessions.json |
Per-provider keys / OAuth sessions (failover + subagents) |
config.toml |
Model, effort, budgets, etc. |
sessions/ |
Chat sessions |
plugins/ · plugins/registry.json |
Marketplace installs (/plugins, nur plugins) |
skills/ · ruflo/ · tool-results/ |
Skills, vector memory, spilled tool output |
status.json · usage.jsonl · ade.json |
Live usage + host panels |
nur.log |
Tracing (not painted into the TUI) |
D. Ecosystem packs
Graphify · PLUR · Ruflo · Executor · agent-browser-cli · Text-to-CAD ·
mobile-harness · /sc-research whitehat DeFi pack (X bounty intel + historical vuln library) · skill packs. Installed when
Node/uv/Bun are available.
Docs: nuroctane.xyz/cli · docs/setup.md
| Surface | What ships |
|---|---|
| TUI | Streaming · duration chips · thought/tool cards · every transcript piece in a theme-coloured frame · peek · drag-select · scrollbar · sessions · multi-provider /login · /model picker · /plugins marketplace · /goal /bro /adhd /scan /btw /codesearch /mc /feedback /tips · every skill as /name · 41 themes including 18 chromatic studies + transparency mode · the window title is the moon alone (phases while a turn runs) · every path in the transcript is a link that opens · block-structured markdown with theme-role colours · tables fit the window, in answers and in tool output · LaTeX rendering (image-peek build) · /receipt verifies the hash chain and exports spans · budgets · doctor |
| Agent | Manual / plan / auto · tools · subagents · todos · auto-compact · session $ / token budgets · Esc cancel · Shift+Tab mid-turn · NL skill auto-activation |
| Vision | look · extract_frames · prompt auto-attach · compact image attachments with on-demand previews (Ctrl+V paste or /image <path>, F4 preview) |
| Tools | read · edit · bash · web · browser (incl. element picking) · git · knowledge · agent · judgments (typesafe) · security agents (enclave) · memory (optmem connectome mem) · diagrams (excalidraw tldraw penecho) · policy (dogwood) · background (bg) · async (admission goal proposal message question) · secrets (secret) · docs (anydoc) · python REPL (repl) · excalidraw |
| Ecosystem | Graphify · GraphJin · PLUR · Ruflo · Executor · TypeSafe/Jev (+ keyless local engines) · Enclave (security agents over native MCP) · OptMem · Headroom · Connectome · dogwood · omp · browser · AKM · 1,000+ skills · /sc-research whitehat DeFi + X bounty intel + historical vuln library · plugin marketplace (Fable, Superpowers, Vercel, …) |
| Hardening | Approval modes · parsed bash denylist · per-domain fetch approval · SSRF blocks · atomic ~/.nur IO · permissions/hooks · signed releases (minisign + SHA-256) · nur doctor |
Why NurCLI
| Real agent, not a wrapper | Modes, tools, streaming, cancel, subagents, auto-compact |
| Sees media | Multimodal images/short video. Sparse frames, not spam. |
| One-shot install | npx nur-cli, one-liner, or Windows EXE · PATH · ecosystem · browser |
| Easy updates | nur update |
| Knowledge stack | Graph · engrams · vector memory · MCP · skills |
| Plugin marketplace | /plugins picker · Superpowers, Fable, Matt Pocock, Addy Osmani, Google, NVIDIA, finance/marketing packs, … → ~/.nur/plugins |
| Natural-language + slash skills | think like fable · TDD this · site cli · HAR file · polish the UI — or /fable-method /adhd /site-cli /<any-skill> |
| Take over other agents | /takeover (alias /hijack) imports Claude Code · Codex · Cursor · Nur · Grok Build sessions |
| Secrets stay local | Keys only in ~/.nur/auth.json (or env) · prefer NUR_API_KEY |
Feature map
Agent harness
- Multi-provider via
/login(72); Responses or Chat Completions adapter - Rotating API keys per provider (
nur auth pool add, or<ENV_KEY>S=k1,k2): a rate-limited, empty or refused key steps aside and the next one takes the request at once - Advisor (
[advisor]): a second model reviews each answer before the turn ends and leaves one note the agent acts on (docs/configuration.md) - Stream rules (
[[stream_rules]],.nur/stream-rules.toml): an answer or tool call that matches a rule's pattern is stopped mid-stream and asked again with the rule in view - Per-directory models (
[[path_models]]): the working directory picks the model, at startup and on/cd - Manual / plan / auto · Shift+Tab mid-turn
- Tool loop, approvals, Esc cancel, subagents, todos, plan mode
- Session budgets (
/budget), tool-result spill, smarter auto-compact - Optional
permissions.toml/hooks.toml /poorcost-saver prompt- Usage ledger:
nur ledger·/ledgertotals seven agent log formats locally, without importing keys or prompt content (docs/ledger.md) /startup-skillroutes to the installed startup validation, competitor research, positioning, and pitch workflows- Project instructions:
AGENTS.md(elseCLAUDE.md) plus aNUR.mdlayer, from the git root down to the working directory, with@imports; user-wide rules in~/.nur/NUR.md(orAGENTS.md) - Session resume:
-c,-r,/sessions(alias/resume) - Takeover cross-agent migration (
/takeover·/hijack): import a session from Claude Code · Codex · Cursor · Grok Build into a native nur session and resume it. Two windows in one modal — presscto switch between/sessionsand/takeover /modelopens a live model list for the active provider (or/model <id>to set one directly)- Modal pickers fit their contents and terminal height, with visible scroll position and consistent keyboard/mouse selection.
/themepreviews align with their click targets; Enter saves and Esc restores the previous palette. /pluginsmarketplace picker (same UX as/login): install Superpowers, Vercel, Firecrawl, Chrome DevTools, Fable, and more into~/.nur/plugins- Natural-language + slash skill activation: plain phrases or
/skill-name//skill-name <prompt>inject the skill body (Fable, TDD, site-cli, adhd, design-eng, resume-*, Excalidraw, …). Palette lists matching skills as you type. Docs /fusion— multi-model debate → one synthesized answer (panel of providers, active model judges)--continuous— sovereign/autonomous mode: loop headless turns toward a goal untilDONEor Ctrl+C/local— run a model locally with bundled llama.cpp (auto-fetchllama-server+ a GGUF sized to your RAM); no API key/bench— benchmark models on your own tasks, replayed in isolated git worktrees and scorednur gateway --chat <id>- run a Telegram bot for one chat; use/inboxand/answerfor questions and required approvals. Without--chatit only pairs; a group also needs--user <id>for the one sender it obeys.nur acp- connect compatible editors over ACP with per-session workspaces, permissions, and stdio MCP servers.
Tools (native)
| Family | Tools |
|---|---|
| read | read_file list_dir sgrep (semantic, Context.dev sgrep, first choice for "where is the code that…") ast_grep (structural, ast-grep: code shapes, rewrite preview and apply) grep glob |
| edit | write_file edit_file multi_edit apply_patch |
| shell | bash |
| vision | look · extract_frames |
| web | web_search web_fetch (treg or TinyFish fallback for blocked or JavaScript-only pages with TREG_TOKEN / TINYFISH_API_KEY) |
| browser | real default browser via agent-browser-cli |
| terminal_browser | in-terminal Chromium (terminal-browser.com); Windows host fallback |
| desktop | native accessibility snapshots and actions through agent-desktop |
| git | git_status git_diff |
| writer isolation | worktree creates a managed Git worktree for an agent child |
| quality | slop_scan (structural slop candidates: god files, duplicated blocks, change coupling, pass-through wrappers) |
| knowledge | graphify graphjin plur ruflo executor skill memory |
| diagrams | excalidraw (hand-drawn .excalidraw via excalidraw-cli) · tldraw (offline desktop boards) · penecho (20k canvas: ink, MathJax, plots) |
| judgments | typesafe (System One typed decisions: gate a call, judge a result, rank, route) |
| security | enclave (Enclave security agents over native MCP: pentests, code security review, findings) · secscan (gitleaks, osv-scanner, semgrep, trivy, checkov and trufflehog when installed; secret values never shown) |
| memory | optmem (permanent, ~/.optmem) · connectome (hierarchical + chronicle) · mem (vector/graph router) · memory · plur ruflo |
| policy | dogwood (Cedar + temporal policy over an agent event stream) |
| context | context (RLM store: register/peek/slice/search) · anydoc (PDF/DOCX/XLSX → markdown) |
| background | bg (jobs that outlive the turn) · admission (async subagent handles) · goal · proposal (staged outputs) · message (peer mail) |
| python | repl (long-lived kernel; state survives compaction) |
| agent | todo_write submit_plan agent · skill harness · codemode (one JS script, several read-only tool calls, one model round) · tool_search (loads deferred tools) · hashline_edit (edits against a tagged read_file snapshot) |
| code intelligence | lsp (language-server queries and post-edit diagnostics) · debug (DAP debugging) · repo_map (ranked file and declaration overview) |
Vision
| Tool | What it does |
|---|---|
look |
Attach workspace images or short video so the model sees them |
extract_frames |
Sparse keyframes via ffmpeg → .nur/frames/<name>/ |
nur "steal UI design tokens from demo.mp4 and scaffold a matching component"
Image attachments: pasted screenshots (Ctrl+V) and /image <path>
stay in a compact composer row until sent. F4 previews or cycles images;
Shift+F4 removes one. Sent attachments offer click-to-peek. Previews render
as real pixels on terminals that support the
kitty graphics protocol,
sixel, or iTerm2 inline images. Queued follow-ups keep their own attachments.
Personal accent colors ([theme] accent = "#..." in config.toml)
layer over any /theme pick. Details: docs/vision.md.
TypeSafe (Jev): a system-one boost layer for whatever provider you are using -
tool gate, result judge, compaction with no summary, skill checks, routing. One
key, every provider. The full Jev library (nur jev models
· /jev models) lists every engine with its native endpoint, local adapter, or
connection recipe; keys are selected per engine. Details:
docs/typesafe.md.
Enclave: Enclave's autonomous security agents (pentests,
code security review, findings, CVE monitoring) through Enclave's MCP server,
spoken natively - no gateway. Enclave is not a chat provider, so its key is a
sidecar like TypeSafe's: nur auth login --provider enclave, then ask for
security work or run /enclave. Details: docs/enclave.md.
NurCLI MCP questions
Which AI providers does NurCLI support?
It supports over 72 providers including OpenAI, Anthropic, Gemini, xAI, Groq, OpenRouter, and local engines like Ollama or LM Studio.
How does NurCLI help reduce token costs?
The harness uses an OMP-style response reserve, avoids duplicate reads, spills large tool results, and supports economy subagents running on the cheapest available models.