Your AI coding assistant — run locally or in the cloud with Ollama.

▶ Watch the full demo with music on YouTube

MandoCode is an AI coding assistant built on RazorConsole, powered by Microsoft Agent Framework and Ollama. RazorConsole makes the entire terminal UI possible — Razor components, a virtual DOM, and Spectre.Console rendering all running in the console.

Run locally or connect to Ollama cloud — no API keys required for anything, including web search (an optional free Tavily key upgrades search reliability). It gives you Claude-Code-style project awareness — reading, writing, searching, planning, and web browsing across your entire codebase — without ever leaving your terminal. It understands any file type: C#, JavaScript, TypeScript, Python, CSS, HTML, JSON, config files, and more.

Prefer a native app? MandoCode Desktop is now available — the exact same engine (Ollama + Microsoft Agent Framework, same brains) with a native Windows interface instead of the terminal: a real integrated shell, a git-aware file explorer, and 16 built-in themes. Same CLI underneath; WinUI 3 on top.

---

Built for the developer workflow

MandoCode organizes your work around agents and workspaces. Each agent has its own model, conversation, settings, and context. Switch between local and cloud Ollama models on the fly, and let agents talk to one another through @ mentions, questions, and delegated tasks.

Save an agent's context as a summary, carry it into another agent, or import multiple snapshots to bring related findings together. Reopen previous agents through history when you are ready to return to their work.

Named workspaces hold up to four independent agents each. Keyboard shortcuts keep agent navigation, file browsing, project directories, context snapshots, MCP servers, and Skills within reach while other agents continue working.

A new way to work in the terminal — v0.16.0

Built on RazorConsole 0.6.0, this release brings multiple workspaces, up to four independent agents per workspace, agent collaboration, portable context snapshots, and saved-agent history into one terminal interface. Explore your files, manage MCP servers and Skills, and keep working while other agents handle their tasks.

The conversation scrolls above a persistent multiline prompt, with searchable pickers and collapsible tool output. See the feature demos below or read the changelog for the full release details.

---

Prerequisites

• .NET 10 SDK: dotnet.microsoft.com/download/dotnet/10.0 (the SDK includes the runtime, so install only the SDK)
• Ollama: ollama.com/download (MandoCode walks you through setup on first run)

Already on .NET 8? You're not blocked. MandoCode ships a .NET 8 build alongside the .NET 10 one, and dotnet tool install -g MandoCode picks the right one for your machine automatically. That fallback goes away when .NET 8 reaches end of support in November 2026, so grab .NET 10 when it's convenient. Run mandocode --doctor to see which runtime you're on.

Install

dotnet tool install -g MandoCode
mandocode

First run launches a guided wizard: it detects Ollama, offers to start it, walks you through cloud sign-in if you'd like more powerful models, and auto-pulls a sensible default. You can re-run it any time with /setup.

Update

Run /update inside MandoCode to check for the latest stable release. For global .NET tool installations, it asks before closing all workspaces, saves agent histories when persistence is enabled, and updates after the CLI exits. Finish work in other agents first. The terminal prints a log path with update progress, success, or failure; launch mandocode again after completion. The .NET SDK is required. Local tools and source builds should use their original installation method.

You can also close MandoCode and update manually:

dotnet tool update -g MandoCode

What Makes MandoCode Different

Multiple Agents and Workspaces Give each task its own agent. Open up to four independent agents per workspace, each with its own model, conversation, and approvals, and organize more projects into named workspaces. Switch between agents while others keep working. Use Alt+N to open an agent, Alt+Left/Right to switch agents, and Alt+Shift+N to create a workspace.	Context You Can Carry Forward Save the useful background from a conversation with /context-snap-create. Capture its decisions, discoveries, and next steps in a summary you can revisit without repeating the whole conversation. Open /context-snap-import (Alt+C) to search saved snapshots, preview a summary, and import it into another agent. Snapshots persist across restarts, making it easy to give a fresh agent a head start.
Return to Previous Work A closed agent does not have to mean a fresh start. Open /history (Alt+H) to find saved agents by name, project, model, or conversation text, and preview their conversations before bringing them back. Re-launch an agent with its name, model settings, transcript, and conversation context—whether you closed it moments ago or are returning to an earlier task.	Agents That Collaborate Bring another agent into the conversation. Type @, then press Tab to switch from Files to Agents. Ask a teammate for its findings, request a review, or hand off independent work while you continue with your task. Expandable agent exchanges let you follow the requests, tool activity, and replies. Each agent keeps its own model, conversation, and approval settings.
Your Project at Your Fingertips Open File Explorer (Alt+E) to browse this agent's project folders and files without leaving the conversation. Select a file to insert an @ reference at the prompt cursor, keeping your existing draft intact. Use Alt+D to change this agent's project directory and Alt+G to review Git changes. Each agent keeps its own project directory, so you can work across projects in the same workspace.
Extend Your Assistant Connect external tools and teach your agents reusable workflows from the CLI. /mcp (Alt+M) opens the MCP manager to add or edit servers, test connections, inspect available tools, and enable or disable integrations. With /skills (Alt+K), install skills from Git, ZIP files, or local folders—or write, generate, and refine your own instructions before saving them. Shared integrations stay available across agents while each agent keeps its own conversation and settings.

---

Features at a Glance

	Feature	Description
Workspace	Independent agents	Up to four agents per workspace, with separate models, conversations, settings, and approvals
Workspace	Named workspaces	Organize projects and switch workspaces while background agents keep working
AI	Agent collaboration	Tag agents with @ to ask questions, request reviews, and hand off independent tasks
Memory	Context snapshots	Save a conversation summary and import it into another agent
Memory	Agent history	Search, preview, and restore closed agents with their conversation context
Tools	File Explorer and Git review	Browse files, attach references, change directories, and inspect project diffs
Tools	Skills manager	Install, edit, generate, refine, and enable reusable skills
Input	Multiline drafts	Wrapped prompts, multiline paste, and cursor navigation before conversation scrolling
Input	Vision attachments	Reference image files; paste clipboard images with Alt+V on Windows
Sharing	HTML transcripts	Export formatted conversations, code, and complete tool output with /transcript-save
AI	Project-aware assistant	Reads, writes, deletes, and searches your entire codebase
AI	Web search & fetch	Web search and webpage reading — keyless via DuckDuckGo, or Tavily with a free API key
AI	MCP server support	Connect to any Model Context Protocol server (stdio or remote HTTP) — Claude-Desktop-compatible config
AI	Session resume	--continue / -c reloads your last conversation for the folder — full memory, tool calls included
AI	Streaming responses	Streams responses to keep long generations alive — no false "stalled" cutoffs
AI	Task planner	Use /plan to review and execute a structured plan
AI	Fallback function parsing	Handles models that output tool calls as raw JSON
UI	Diff approvals	Color-coded diffs with approve / deny / redirect
UI	Markdown rendering	Rich terminal output — headers, tables, code blocks, quotes
UI	Syntax highlighting	C#, Python, JavaScript/TypeScript, Bash, PowerShell
UI	Clickable file links	OSC 8 hyperlinks for file paths
UI	Terminal theme detection	Auto-adapts colors for light and dark terminals
UI	Taskbar progress	Windows Terminal integration during task execution
Input	/ command autocomplete	Slash commands with dropdown navigation
Input	@ file references	Attach file content to any prompt
Input	! shell escape	Run shell commands inline (!git status, !ls)
Input	/copy and /copy-code	Copy responses or code blocks to clipboard
Music	Lofi + synthwave	Bundled tracks with playback, volume, and genre controls
Config	Configuration wizard	Guided setup with model selection and connection testing
Config	Config validation	Auto-clamps invalid settings to safe ranges
Reliability	Retry + deduplication	Exponential backoff and duplicate call prevention
Education	/learn command	LLM education guide with optional AI educator chat

---

Try it: carry a task between agents

1. Open MandoCode in your project and ask it to explain a file using an @ reference.
2. Press Alt+N to open another agent. Use Alt+Left/Right to move between them; each keeps its own conversation.
3. In a prompt, type @, press Tab, and select another agent. Ask it to share its findings or review your changes.
4. Run /context-snap-create in the agent with useful background. Switch to the receiving agent, run /context-snap-import, choose the summary, and send your next prompt with that context.
5. Save a shareable conversation with /transcript-save. After closing an agent, use /history to preview and restore it later.

Press Alt+Shift+N when you want another workspace for a different project or task. Use /keybindings for the full shortcut reference.

---

Everyday keyboard shortcuts

Shortcut	Action
Alt+N	Open a new agent
Alt+Left / Right	Switch agents
Alt+W	Close the highlighted agent, including from an open menu; stop active work if needed
Alt+Shift+N	Create a workspace
Alt+Shift+Left / Right	Switch workspaces
Alt+Shift+A	Open the workspace picker
Alt+Shift+W	Close the current workspace when its agents are idle
Alt+S	Open this agent's settings
Alt+C	Open context snapshots
Alt+H	Browse saved-agent history
Alt+E	Open File Explorer
Alt+D	Change this agent's project directory
Alt+G	Review Git changes
Alt+M	Open the MCP manager
Alt+K	Open the Skills manager
Alt+V	Attach a clipboard image (Windows, vision-capable models)
Tab in the @ picker	Switch between Files and Agents
Page Up / Page Down	Scroll the conversation from the prompt
Ctrl+End	Return to the latest conversation output

Use /keybindings for the full reference, including menu navigation and editor controls. Type / to search available commands and see their descriptions. Shortcuts depend on your terminal forwarding the key combination to MandoCode.

---

Commands

Type / to see the autocomplete dropdown, or ! to run a shell command.

Agents and workspaces

Command	What it does
/agent-new	Open an independent agent pane (Alt+N)
/agent-focus	Switch agents; optional left/right (Alt+Left/Right)
/agent-close	Close the current idle agent
/agent-rename <name>	Rename this agent
/agent-settings	Edit this agent's model, behavior, and integrations (Alt+S)
/agent-file-explorer	Browse this agent's files and folders (Alt+E)
/workspace-new [name]	Create a workspace with up to four agents
/workspace [name or number]	Switch workspaces
/workspace-all	Search open workspaces
/workspace-rename <name>	Rename this workspace
/workspace-close	Close this workspace when its agents are idle

Conversation, context, and sharing

Command	What it does
/plan <goal>	Generate a plan for review before execution
/plan-resume	Continue an unfinished plan
/plan-discard	Forget an unfinished plan
/context-snap-create	Save a context snapshot (summary) of this agent conversation (Alt+C)
/context-snap-import	Browse and import saved context snapshots (Alt+C)
/copy	Copy last AI response to clipboard
/copy-code	Copy code blocks from last response
/transcript-save	Save this agent conversation as standalone HTML; optional quoted file path
/clear	Clear conversation history
/history	Search and restore closed agents, including transcripts and conversation context (Alt+H)

Project tools and integrations

Command	What it does
/git-changes	Review this agent's changed files and Git diffs (Alt+G)
/change-directory [path]	Change this agent's project directory (Alt+D)
/command <cmd>	Run a shell command
/mcp	Open the shared MCP manager: add/edit, test, enable/disable, and inspect tools (Alt+M)
/mcp tools <server>	List tools exposed by connected MCP servers (server optional)
/mcp-reload	Restart all MCP servers and re-register their tools
/skills	Manage user skills: install from Git/ZIP/folder, edit, generate/refine, and enable/disable (Alt+K)
!<cmd>	Shell escape (e.g., !git status)
!cd <path>	Change project root directory

Setup, models, and help

Command	What it does
/help	Show commands and usage examples
/keybindings	Show keyboard shortcuts grouped by task
/update	Check for a stable release and update after the CLI exits
/setup	Guided wizard — reconnect to Ollama, install/sign in, or pick a different model
/model	Quick switch — pick a different model (context window auto-sized for local tiers)
/config	Adjust settings — guided wizard
/config set <key> <value>	Set one setting inline without leaving the session (e.g. /config set modelResponseTimeout 300); no args lists all keys + current values
/retry	Retry Ollama connection
/ollama-pull	Browse the Ollama library and download a model
/ollama-serve	Start local Ollama if unavailable and retry the connection twice
/learn	Interactive guide to LLMs and local AI
/exit	Exit MandoCode

Music

Command	What it does
/music	Start playing music
/music-stop	Stop playback
/music-pause	Pause / resume
/music-next	Next track
/music-vol <0-100>	Set volume
/music-lofi	Switch to lofi
/music-synthwave	Switch to synthwave
/music-list	List available tracks

Downloading Ollama models

First-run setup and /setup use a guided panel: Connect → Choose model → Download → Verify → Ready. Choose a recommended cloud starter (glm-5.3-flash:cloud or deepseek-v4.1-flash:cloud), a local starter with approximate download sizes, an installed model, or the full model browser. Cloud requires Ollama sign-in and internet; account limits and pricing apply. Local model memory use depends on the model and context, beyond download size. Back returns from choices, and Escape pauses setup. Downloads target your configured Ollama server; remote cloud authentication must be configured on that server. Setup tests a real model response before saving it as the default for new agents. Failed verification offers retry, another model, or Finish later. Existing model preferences are preserved until verification succeeds.

Use /ollama-pull anytime to browse the Ollama library, choose a model tag, and download it with progress. Use the search box to filter model names; choose Popular, Name A–Z, or Newest (the default) (Ctrl+S cycles sorting). The table shows model names, capabilities, and hosting. Toggle Cloud to show only models marked for cloud hosting. Tab cycles through search, sorting/filter buttons, and the list. Left/Right selects a sorting button; Enter applies it. Ctrl+P pulls the model name typed in the search field. Enter, Escape, or the Cancel button cancels a download. Use /model afterwards to switch to the downloaded model.

Mention another agent

Type @ for files, then Tab to switch to open agents across workspaces. Type a name to filter, use Up/Down and Enter to select, or press Tab again for files. The switching hint appears only in the picker title. Escape closes the picker without deleting your text. Names with spaces are quoted automatically. Selecting a file that shares an agent's name inserts an explicit @./ path.

Try What is @Ares working on? or Ask @Ares what it found and let me know later. The picker shows plain agent names in purple, with white text when highlighted. Recognized chat mentions appear with consistent agent avatars and purple names in submitted prompts and reply prose. File paths, email addresses, code snippets, and unknown names stay literal. Agent interactions have their own collapsed sections in both transcripts, showing participants and live status. Expand one to inspect requests, messages, task commentary, tool calls, and Markdown replies with timestamps. Expansion stays open as updates arrive; failures and cancellations also have a visible notice. The main answer to the user stays outside the section. Up/Down moves through multiline drafts first, then scrolls the conversation when the caret reaches its first or last line. Each agent uses its own model, context, tools, and approval settings; busy agents keep working and call loops are blocked.

Agent tool	Behavior
ask_agent_and_wait	Waits for an answer needed to continue the current request
ask_agent_async	Asks an independent question; its answer arrives later
send_agent_message	Delivers information to an inbox without starting a turn or requesting a reply
delegate_to_agent	Assigns independent work and returns after acceptance
request_agent_review	Assigns a background review; file changes, commands, plans, and further assignments are blocked
handoff_to_agent	Transfers a clearly specified task scope; the caller is instructed to stop working on that scope
update_agent_job	Queues additional context for an interaction by ID
cancel_agent_job	Requests cancellation of an interaction you assigned; completed changes remain
wait_for_agent_job	Waits explicitly for a result, with a timeout; ending the wait does not cancel the job
list_agents, get_agent_status, read_agent_transcript, check_delegations	Inspect agents, conversations, and background interaction states

Background questions and jobs have readable IDs such as job1; inbox messages use message1. Jobs and report Completed, Declined, Cancelled, or Failed outcomes. They return after acceptance rather than waiting for a result. Reports appear in the caller's transcript and are included with its next request. Messages and updates can reach busy agents but are consumed at the next turn, not injected into an active request. Cancellation targets only that interaction; ownership is checked by the assigning agent. Handoffs communicate responsibility and file scope to the models rather than locking files. Reviews permit built-in inspection tools; shell commands and unknown MCP tools are unavailable in a read-only review.

Context snapshots

Open Context Snapshots beside Agent Settings, or use /context-snap-create to save a conversation summary. The outlined name textbox is ready to type into as soon as the form opens; no Enter-to-edit step is needed. Tab moves to the model and buttons. Enter a name or leave it blank for an AI title, choose a summarizer model (initially the current agent model), then Create. The separate summarizer does not change your agent context or run tools. Saved snapshots persist in ~/.mandocode/snapshots.json and are shared across CLI agents and workspaces. The browser has a search/Create/Close toolbar, collapsible project groups with newest snapshots first, and a summary pane with Import into Agent/Delete actions. Wide terminals show the list and summary side by side; narrow terminals stack them. Tab moves through search, list, toolbar, summary, and actions. The selected snapshot stays subtly shaded while another control has keyboard focus. Click project headings or press Enter on them to collapse/expand. Click the summary to focus it; mouse wheel, arrows, Home/End, and Page Up/Page Down scroll it. Enter on a snapshot imports it, while Enter on Close or Escape closes without importing. Import queues a summary for the next message; multiple imports stack without duplicating the same snapshot. Delete requires confirmation. When /model, /setup, or the configuration wizard clears an existing conversation, a prompt offers Keep Memory, Create Snapshot, or Dismiss. Changing models in Agent Settings already keeps memory.

MCP and Skills managers

Open /mcp (Alt+M) for MCP Servers or /skills (Alt+K) for Skills, or choose Manage MCP Servers / Manage Skills in Agent Settings → Integrations. Each table row has an Active/Disabled status toggle, Edit, and Delete buttons; MCP rows also have Tools. Click an item's name or press Enter on it to view full details. Up/Down moves between rows, Left/Right selects row actions, Tab visits all controls, and Escape returns or closes. Buttons wrap beneath their item in narrow panes. New Skill offers Write Manually with field explanations, or Create with AI with a task description and model picker. Names, URLs, commands, skill descriptions, and AI requests use outlined textboxes you can type into directly. Tab or Up/Down moves between fields; Left/Right moves the text cursor. Instructions, arguments, headers, and environment variables open dedicated multiline editors. Manual forms contain only Name, Description, and Instructions; AI refinement is optional and opens separately. Generated drafts remain editable until Save. Install offers a folder browser or Git URL/ZIP input. The MCP table has a live search textbox with +MCP Server to its right; Down or Enter from search moves into the results. Skills has a focused Search Skills textbox beside +Skill and Install. Both search fields use gold text and filter as you type. The selected item has its full name and wrapped details directly below the table; bulk toggles, Refresh, Close, and operation feedback stay in the footer. These managers edit shared user settings; Agent Settings still controls whether a particular agent can use MCP tools.

For MCP, choose +MCP Server, select stdio or HTTP, fill in the executable/URL, then Test Connection or Save & Connect. Enter stdio arguments one per line, keeping arguments with spaces together without extra quotes. Environment variables and HTTP headers use one KEY=value per line. Enable/Disable and Enable All/Disable All persist the selection and reconnect open agents without clearing their conversations. Connection failures remain visible in the list; Details shows the complete error. Installing a stdio server means configuring its launch command, such as npx or uvx; the executable must be available on your system.

For Skills, Install accepts a Git repository URL, local ZIP archive, or folder containing one or more SKILL.md directories. It copies the skill's bundled files, skips existing names, and does not alter project skills. +Skill offers Write Manually or Create with AI. Manual creation and Edit provide name, description, and instructions fields; AI creation takes a task description and authoring model before generating a draft. Review the Instructions before Save. Form actions are Save, Cancel, and Refine with AI. Cancel immediately discards the draft and returns to the list, as it does when creating or editing MCP servers. Authoring uses a separate conversation with no tools and leaves the agent's conversation and model unchanged.

Skills use the same enable/disable convention as Desktop (SKILL.md / SKILL.md.disabled). Removing a skill moves it into .trash inside your user skills directory, with its recovery path shown in Details. Shared changes wait until other agents finish their requests and close their settings menus. Project skills keep their existing precedence over user skills.

Setup vs config vs model

• /setup — first-run wizard, guided. Detects Ollama, offers to install it, walks you through cloud sign-in, picks a model with hardware-aware tiers, auto-pulls a sensible default. Use when something's broken or you're a newcomer.
• /model — quick switch. Pick a model from your pulled list and go — local picks get a context window sized to their hardware tier automatically. Use when you just want to swap models.
• /config — adjust settings. Full configuration form covering temperature, timeouts, ignore dirs, etc. Use when you know exactly what knob you want to turn.

CLI flags (outside the chat loop)

mandocode --continue        # resume this folder's most recent conversation (alias: -c)
mandocode --doctor          # preflight check: .NET runtime, Ollama status, models, sign-in
mandocode --config show     # print current config
mandocode --config init     # create a default config file
mandocode --config set <key> <value>   # set a single value (e.g. set model qwen3:8b)
mandocode --config path     # show config file location

Run mandocode --doctor any time chat is misbehaving — exits 0 if everything's green, 1 if anything's missing, with a clear summary of what's wrong.

Conversations save themselves automatically at the end of every turn (per project folder, under ~/.mandocode/sessions/), so mandocode --continue picks up exactly where you left off — the model remembers what it read, said, and did, tools included. /clear deletes the saved session too.

---

How It Works

  You type a prompt
        |
  MandoCode adds project context (@files, system prompt)
        |
  Microsoft Agent Framework sends to Ollama (local or cloud model)
        |
  AI responds with text + function calls
        |
  File operations go through diff approval
  Web searches and fetches run directly
        |
  Rich markdown rendered in your terminal

The AI has sandboxed access to your project through a FileSystemPlugin (9 functions: list files, glob search, read, write, delete files/folders, text search, path resolution) and a WebSearchPlugin (web search via Tavily or DuckDuckGo, webpage fetching — works without any API key). All file operations are locked to your project root — path traversal is blocked.

---

Troubleshooting

mandocode --doctor

Prints your runtime version, Ollama status, models pulled, and cloud sign-in state.

Local models: the context window is managed for you

Using cloud models (:cloud tags)? Skip this section. Cloud model context is managed on Ollama's servers and set to the model's maximum by default — nothing on your machine affects it, including the desktop app's slider.

The context window is how much conversation + code the model can see at once. Left to its own devices, Ollama defaults it to ~4k tokens — which an agentic session fills almost immediately — and never errors on overflow: it silently drops the oldest content (including the system prompt, the model's instructions!), which looks like the model suddenly getting dumber mid-conversation. MandoCode handles all of this:

• Auto-sized per model — picking a local model in /setup or /model sizes the window to its hardware tier: 16k under 7B, 32k for 7B+ (bigger models imply bigger GPUs, which also cover the larger KV cache). The floor is 16k because MandoCode's system prompt and tool definitions consume most of an 8k window before the conversation even starts.
• Enforced on every request — the window is sent as num_ctx with each chat call, which outranks the Ollama desktop app's slider, OLLAMA_CONTEXT_LENGTH, and the ~4k default. Changes apply from your next message; no daemon restarts.
• Overflow protection — if a long conversation nears the window anyway, MandoCode compacts older history into a recap before sending, instead of letting Ollama truncate silently.

When to tune it yourself (mandocode --config set contextLength 16384, live from the next message):

• Generation got slow after switching models — the auto-picked window may not fit your GPU. Every 8k of window costs roughly 0.5–1.5 GB of VRAM depending on the model; when the KV cache spills out of VRAM, tokens/sec craters. (Seen in the wild: a 256k window dropped a small model from ~175 tok/s to ~11.) More window is not better — step it down a notch.
• Long sessions with lots of tools (MCP servers, web search) — tool definitions ride on every request and can crowd even a 16k window. Step up to 32k if you have the memory.
• You'd rather manage it from Ollama — set it to 0 and the daemon's own setting governs (desktop app: Settings → Context length).

Verify what a loaded model is actually using with ollama ps (look at the CONTEXT column). Run /learn inside MandoCode for a friendly explainer.

⚠️ All models: check your response cap (the #2 gotcha)

The context window's evil twin — and unlike the slider above, this one applies to every model, cloud included. If the model announces work and then just stops — "I'll create the game…" and the turn ends with no plan, no files, and no error — your maxTokens is too low. It caps a single reply (NumPredict), and reasoning models spend output tokens thinking before they emit a tool call, so a low cap cuts them off before they ever act.

Fresh installs default to 32k and never notice it. But if your config predates v0.11, or you once lowered maxTokens thinking it was the context window (they're different knobs — this caps what the model says, the context window caps what it sees), check it:

mandocode --config show                  # look at "Max Tokens"
mandocode --config set maxTokens 32768

The telltale sign: token tracking shows output pinned at exactly your cap, turn after turn (e.g. 2k out every time). Note that a running session keeps the config it loaded at startup — restart MandoCode (or use /config set in-app) for the change to take effect.

Build from source

git clone https://github.com/DevMando/MandoCode.git
cd MandoCode
dotnet build src/MandoCode/MandoCode.csproj
dotnet run --project src/MandoCode/MandoCode.csproj -- /path/to/your/project

---

Recommended Models

Models with tool/function calling support work best with MandoCode. The first-run wizard offers exactly the models below — auto-pulls the cloud default, or lets you pick a local tier matched to your hardware.

Cloud (no GPU required — runs on Ollama's servers; sign in with ollama signin. Account limits and pricing apply):

Model	Notes
glm-5.3-flash:cloud	Default cloud starter — offered by /setup alongside deepseek-v4.1-flash:cloud
minimax-m3:cloud	General-purpose alternative
kimi-k2.7-code:cloud	Code-focused

Local (fully offline, runs on your hardware):

Model	Size	Notes
qwen3:8b	~5.2 GB	Recommended — best balance of speed & quality (~6 GB VRAM)
qwen2.5-coder:7b	~4.7 GB	Code-focused (~5–6 GB VRAM)
mistral	~4.1 GB	Fast, general-purpose 7B (~5 GB VRAM)
llama3.1	~4.9 GB	Strong general reasoning, 8B (~6 GB VRAM)

MandoCode validates model compatibility on startup. Run /learn for a detailed guide on model sizes and hardware requirements, or /setup to switch between tiers any time.

---

Configuration Reference

Config File

Located at ~/.mandocode/config.json

{
  "ollamaEndpoint": "http://localhost:11434",
  "modelName": "glm-5.3-flash:cloud",
  "modelPath": null,
  "temperature": 0.7,
  "maxTokens": 32768,
  "contextLength": 16384,
  "requestTimeoutMinutes": 15,
  "modelResponseTimeoutSeconds": 680,
  "responseStreaming": "all",
  "toolResultCharBudget": 100000,
  "enableAutoContinuation": true,
  "maxAutoContinuations": 3,
  "markdownRenderTimeoutSeconds": 60,
  "ignoreDirectories": [],
  "enableDiffApprovals": true,
  "enableTaskPlanning": true,
  "enableTokenTracking": true,
  "enableThemeCustomization": true,
  "enableFallbackFunctionParsing": true,
  "enableWebSearch": true,
  "tavilyApiKey": null,
  "enableMcp": true,
  "mcpServers": {},
  "functionDeduplicationWindowSeconds": 5,
  "maxRetryAttempts": 2,
  "music": {
    "volume": 0.5,
    "genre": "lofi",
    "autoPlay": false
  }
}

All Options

Key	Default	Description
ollamaEndpoint	http://localhost:11434	Ollama server URL
modelName	glm-5.3-flash:cloud	Model to use
modelPath	null	Optional path to a local GGUF model file
temperature	0.7	Response creativity (0.0 = focused, 1.0 = creative)
maxTokens	32768	Cap on a single reply (NumPredict) — a runaway-generation safety ceiling, not the context window. If the model announces work then stops without acting, this is too low (see Troubleshooting)
contextLength	16384	Context window (num_ctx / KV-cache size) for local models — sent with every request, so it overrides the Ollama desktop app's slider and the daemon's ~4k default; auto-sized to the model's hardware tier by /setup and /model. 0 = defer to the daemon's own setting. Bigger window = more VRAM. Cloud models manage context server-side
requestTimeoutMinutes	15	Per-turn ceiling for a single model call / plan step. Cancel anytime with Ctrl+C
modelResponseTimeoutSeconds	680	Stall watchdog — max seconds a single model call may run before it's treated as stalled. With streaming on (the default), the per-chunk heartbeat makes this mostly a safety net
responseStreaming	all	Which models stream (with a per-chunk stall-watchdog heartbeat): all, cloud (cloud only), or off (non-streaming everywhere). true/false are accepted as aliases for all/off
toolResultCharBudget	100000	Total characters of tool results per turn/step before further tool calls are refused (≈25k tokens) — guards the context window
enableAutoContinuation	true	When the tool-result budget is exhausted mid-task, auto-continue in a fresh scope instead of stopping
maxAutoContinuations	3	Hard cap on auto-continuations per request (prevents runaway loops)
markdownRenderTimeoutSeconds	60	Max seconds to render the final markdown before falling back to raw text
ignoreDirectories	[]	Additional directories to exclude from file scanning
enableDiffApprovals	true	Show diffs and prompt for approval before file writes/deletes
enableTaskPlanning	true	Enable task-planning support; start a CLI plan explicitly with /plan
enableTokenTracking	true	Show session token totals and per-response token costs
enableThemeCustomization	true	Detect terminal theme and apply a curated ANSI palette
enableFallbackFunctionParsing	true	Parse function calls from text output
enableWebSearch	true	Enable web search & page fetch (DuckDuckGo keyless, or Tavily with a key)
tavilyApiKey	null	Optional Tavily key for reliable AI-optimized search (or set the TAVILY_API_KEY env var)
enableMcp	true	Enable MCP server integration (servers configured under mcpServers)
functionDeduplicationWindowSeconds	5	Time window to prevent duplicate function calls
maxRetryAttempts	2	Max retry attempts for transient errors
music.volume	0.5	Music volume (0.0 - 1.0)
music.genre	lofi	Default genre (lofi or synthwave)
music.autoPlay	false	Auto-start music on launch

CLI Config Commands

mandocode config show              # Display current configuration
mandocode config init              # Create default configuration file
mandocode config set <key> <value> # Set a configuration value
mandocode config path              # Show configuration file location
mandocode config --help            # Show help

Environment Variables

Variable	Overrides
OLLAMA_ENDPOINT	ollamaEndpoint in config
OLLAMA_MODEL	modelName in config

Diff Approvals — Deep Dive

When the AI writes or deletes a file, MandoCode intercepts the operation and shows a color-coded diff before applying changes.

What You See

• Red lines — content being removed
• Light blue lines — content being added
• Dim lines — unchanged context (3 lines around each change)
• Long unchanged sections are collapsed with a summary

Approval Options

Option	Behavior
Approve	Apply this change
Approve - Don't ask again	Auto-approve future changes to this file (per-file), or all files (global)
Deny	Reject the change, the AI is told it was denied
Provide new instructions	Redirect the AI with custom feedback

For new files, "don't ask again" sets a global bypass — all future writes and deletes are auto-approved for the session. For existing files, the bypass is per-file.

Even when auto-approved, diffs are still rendered so you can follow along.

Delete Approvals

File deletions show all existing content as red removals with a deletion warning. The same approval options apply.

Toggle

mandocode config set diffApprovals false

@ File References — Deep Dive

Type @ anywhere in your input (after a space or at position 0) to trigger file autocomplete. A dropdown appears showing your project files, filtered as you type.

How It Works

1. Type your prompt and hit @ — a file dropdown appears
2. Type a partial name to filter (e.g., Conf) — matches narrow down
3. Use arrow keys to navigate, Enter to select; Tab switches between Files and Agents
4. The selected path is inserted (e.g., @src/MandoCode/Models/MandoCodeConfig.cs)
5. Continue typing and press Enter to submit
6. MandoCode reads the referenced file(s) and injects the content as context for the AI

Examples

explain @src/MandoCode/Services/AIService.cs to me
what does the ProcessFileReferences method do in @src/MandoCode/Components/App.razor
refactor @src/MandoCode/Models/LoadingMessages.cs to use fewer spinners

Multiple @ references in one prompt are supported. Files over 10,000 characters are automatically truncated.

Controls

Key	Action
@	Open file dropdown
Type	Filter files by name
Up/Down	Navigate dropdown
Enter	Insert selected file path (does not submit)
Tab	Switch between Files and Agents
Escape	Close dropdown, keep text
Backspace	Re-filter, or close if you delete past @

Task Planner — Deep Dive

Start with /plan <goal> when you want to review a structured plan before work begins. Ordinary CLI prompts run directly; complexity alone does not switch them into plan mode.

Workflow

1. Plan generation — describe the goal and review the proposed steps.
2. User approval — approve the plan, edit steps, choose One-shot it, or cancel.
3. Execution — follow step-by-step progress and approve changes as needed.
4. Recovery — review available recovery choices if a step fails; use /plan-resume to continue an unfinished plan or /plan-discard to forget it.

See Task Planner Documentation for full technical details.

/learn — LLM Education

The /learn command helps new users understand local LLMs and get set up.

Scenario	What happens
Startup, no Ollama detected	Automatically displays the educational guide instead of a bare error
/learn typed, no model running	Displays the static educational guide
/learn typed, model is running	Shows the guide, then offers to enter AI educator chat mode

Educational Content

1. What are Open-Weight LLMs? — Free, private, offline models vs. cloud AI
2. Model Sizes & Hardware — Parameters, quantization, VRAM requirements
3. Cloud vs Local Models — Ollama cloud models (no GPU) vs local models
4. Recommended Models — Table of cloud and local options
5. Getting Started — Step-by-step setup instructions

AI Educator Chat Mode

When Ollama is running, /learn offers an interactive chat mode where the AI explains LLM concepts using beginner-friendly language. Type /clear to return to normal mode.

MCP Servers — Deep Dive

MandoCode speaks the Model Context Protocol as a client, which means you can plug in any published MCP server — filesystem, database, GitHub, Linear, Slack, whatever — and its tools show up to the model alongside MandoCode's built-in plugins.

Adding a server

Two ways:

• /mcp inside MandoCode → Add Server — the server editor for name, transport, URL/command, and optional headers/env vars, with Test Connection and Save & Connect.
• Hand-edit ~/.mandocode/config.json — useful when copy-pasting a mcpServers block from a server's README. Run /mcp-reload after saving.

Config shape

The mcpServers block mirrors Claude Desktop's schema, so you can copy-paste any server's README installation snippet directly into ~/.mandocode/config.json:

{
  "enableMcp": true,
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allow"]
    },
    "solana": {
      "url": "https://mcp.solana.com/mcp",
      "transport": "http"
    },
    "github": {
      "url": "https://api.githubcopilot.com/mcp/",
      "headers": { "Authorization": "Bearer ghp_your_token_here" },
      "autoApprove": ["list_issues", "get_pr"]
    }
  }
}

Transports

• stdio — for local servers. Populate command + args + optional env. Works with any server published as an npm/pip/go binary.
• HTTP / SSE — for remote servers. Populate url; the client auto-detects Streamable HTTP or SSE. Custom headers go in headers — most commonly Authorization: Bearer … for servers that accept static tokens.

Does MandoCode need Node?

No. MandoCode itself is pure .NET. But individual servers may need whatever runtime their command points at — Node for npx, Python for uvx, or nothing extra for standalone binaries. Same situation as Claude Desktop, Cursor, and VS Code.

OAuth-only servers

Native OAuth is not in this release. For servers that require an OAuth flow (some hosted connectors like Google Drive), wrap them in stdio via the community mcp-remote proxy, which handles the browser dance itself:

"gdrive": {
  "command": "npx",
  "args": ["mcp-remote", "https://example.com/mcp"]
}

Approvals

MandoCode cannot tell a read-only MCP tool from a destructive one by inspecting arguments, so the first call of each (server, tool) pair prompts you with Approve / Approve for session / Deny. Pre-trusted tools can be listed under autoApprove in a server's config entry to skip the prompt entirely.

Slash commands

• /mcp — opens the shared server manager with connection status and tool inspection
• /mcp → Add Server — configure a new server without hand-editing JSON
• /mcp → Delete on a server row — remove a server from config (with confirmation)
• /mcp tools <server> — list every tool exposed by connected servers with descriptions (server arg optional — omit to list all)
• /mcp-reload — tears down every MCP client, restarts them, and re-registers their tools on the agent (useful when you edit the config mid-session)

Toggle

mandocode --config set mcp false   # disable all MCP integration

Individual servers can be muted without deleting them — set "disabled": true on any entry in mcpServers.

AI Plugins

FileSystemPlugin

The AI has sandboxed access to your project directory through these functions:

Function	Description
list_all_project_files()	Recursively lists all project files, excluding ignored directories
list_files_match_glob_pattern(pattern)	Lists files matching a glob pattern (*.cs, src/**/*.ts)
read_file_contents(relativePath, startLine?, endLine?)	Reads file content with line count — large files page via startLine/endLine, and truncated output names the exact line to resume from
write_file(relativePath, content)	Writes/creates a file (creates directories as needed)
delete_file(relativePath)	Deletes a file
create_folder(relativePath)	Creates a new directory
delete_folder(relativePath)	Deletes a directory and all its contents
search_text_in_files(pattern, searchText)	Searches file contents for text, returns paths and line numbers
get_absolute_path(relativePath)	Converts a relative path to absolute

Security: All operations are sandboxed to the project root. Path traversal is blocked with a separator-boundary check.

Ignored directories: .git, node_modules, bin, obj, .vs, .vscode, packages, dist, build, __pycache__, .idea — plus any custom directories from your config.

WebSearchPlugin

The AI can search the web and fetch page content — no API keys required.

Function	Description
search_web(query, maxResults)	Searches the web and returns titles, URLs, and snippets (1–10 results)
fetch_webpage(url, maxCharacters)	Fetches a URL and extracts readable text content (500–15,000 chars)

Out of the box, search uses DuckDuckGo's free HTML endpoint — which rate-limits and temporarily blocks IPs under heavy agentic use, so searches can randomly fail. For reliable, AI-optimized search, add a free Tavily API key (free tier ~1,000 searches/month):

/config set tavilyKey tvly-...        # in-app — verifies the key live against Tavily
mandocode --config set tavilyKey tvly-...   # or from the CLI

With a key set, search_web prefers Tavily and keeps DuckDuckGo as the fallback; clear it anytime with /config set tavilyKey clear. The key is stored locally in ~/.mandocode/config.json and only ever sent to Tavily — set the TAVILY_API_KEY environment variable instead if you'd rather keep it out of the file. Fetched pages are cleaned of scripts, nav, and non-content elements via HtmlAgilityPack.

Reliability & Internals

Retry Policy

Transient errors (HTTP failures, timeouts, socket errors) are retried with exponential backoff:

Attempt 1 -> fail -> wait 500ms
Attempt 2 -> fail -> wait 1000ms
Attempt 3 -> fail -> throw

Function Deduplication

Operation	Window	Matching
Read operations	2 seconds	Function name + arguments
Write operations	5 seconds (configurable)	Function name + path + content hash (SHA256)

Fallback Function Parsing

Some local models output function calls as JSON text instead of proper tool calls. MandoCode detects and parses:

• Standard: {"name": "func", "parameters": {...}}
• OpenAI-style: {"function_call": {"name": "func", "arguments": {...}}}
• Tool calls: {"tool_calls": [{"function": {"name": "func", "arguments": {...}}}]}

Markdown Rendering

AI responses are rendered as rich terminal output:

Markdown	Rendered as
**bold**	Bold text
*italic*	Italic text
`code`	Cyan highlighted
Fenced code blocks	Bordered panels with syntax highlighting
Tables	Spectre.Console table widgets
# Headers	Bold yellow with horizontal rules
- lists	Indented bullet points
> quotes	Grey-bordered block quotes
URLs	Clickable OSC 8 hyperlinks

Syntax highlighting supports C#, Python, JavaScript/TypeScript, and Bash with language-specific keyword coloring.

Token Tracking

• Per-response: [~1.2k in, 847 out] after each AI response
• Session total: Total [4.2k tokens] above the prompt
• File estimates: @file attachments show estimated token cost (chars/4)

Event-Based Completion Tracking

Function executions use semaphore-based signaling, ensuring each task plan step fully completes before the next begins.

---

Developer check: updater simulation

Developers can exercise the real updater helper without installing or updating any tool:

dotnet run --project tests/UpdateSimulation/UpdateSimulation.csproj

This simulation replaces the package update command, verifies the helper waits for its parent to exit, and runs success and failure cases. Logs remain under bin/update-simulation. It does not test an actual NuGet installation or the interactive confirmation/history workflow.

Architecture

src/MandoCode/
  Components/        Razor UI (App, Banner, HelpDisplay, ConfigMenu, Prompt)
  Services/          Core logic (AI, markdown, syntax, tokens, music, diffs, input state machine)
  Models/            Data models, config, system prompts, educational content
  Plugins/           Tool classes bound as AIFunctions (FileSystem, WebSearch, Planning, Skills)
  Audio/             Bundled lofi and synthwave MP3 tracks
  docs/              Feature and architecture documentation
  Program.cs         Entry point and DI registration

Dependencies

Package	Purpose
Microsoft.Agents.AI 1.19.0	Agentic orchestration — tool calling, approval middleware
Microsoft.Extensions.AI 10.9.0	Provider-agnostic chat/tool abstractions underneath Agent Framework
OllamaSharp 5.4.30	Ollama model integration
ModelContextProtocol 2.2.0	MCP server support
RazorConsole.Core 0.6.0	Terminal UI with Razor components
Markdig 1.3.2	Markdown parsing
NAudio 2.2.1	Audio playback
HtmlAgilityPack 1.12.4	HTML parsing for web search
FileSystemGlobbing 10.0.11	Glob pattern matching
YamlDotNet 18.1.0	Skill/config YAML parsing

---

Why .NET?

Most AI coding agents in the wild are built with Python, Rust, or TypeScript. .NET rarely gets mentioned — but it should.

Microsoft Agent Framework is Microsoft's open-source SDK for building AI agents, and it's one of the most capable orchestration frameworks available: tool/function calling, middleware you can hook into for approvals and circuit breakers, structured multi-agent workflows, and first-class support for local models through providers like Ollama. It runs cross-platform on Windows, Linux, and macOS.

MandoCode exists partly to prove the point: you can build a full-featured, agentic CLI tool on .NET and Agent Framework that stands alongside anything built in other ecosystems. The tooling is there. It's open source. It just doesn't get the attention it deserves.

---

MIT License

---

Support MandoCode

If MandoCode helps your workflow, you can support its development by buying me a coffee.