MandoCode

MandoCode is a local-first CLI coding agent for developer workflows, written in .NET C#, and technically powered by RazorConsole for the terminal UI (Razor components, virtual DOM, and Spectre.Console rendering), and driven by Microsoft Agent Framework and Ollama. It can run locally or connect to Ollama Cloud; it is officially said to require no API keys of any kind, including for web search; an optional free Tavily key can improve search reliability. It provides Claude Code-like project awareness, capable of reading, writing, searching, planning, and web browsing across the entire codebase within the terminal, and understands multiple file types such as C#, JavaScript, TypeScript, Python, CSS, HTML, JSON, configuration files, and more.

The core organization revolves around "Agents" and "Workspaces." Each agent has its own model, conversation, settings, and context; you can instantly switch between local and cloud Ollama models; agents can collaborate through @ mentions, asking questions, and delegating tasks. Each workspace holds up to four independent agents. You can save an agent's context as a summary snapshot, bring it into another agent, or import multiple snapshots to consolidate related findings. You can also reopen previous agents through history. Named workspaces can each hold up to four independent agents, keyboard shortcuts support agent navigation, file browsing, project directories, context snapshots, MCP servers, and Skills management, while other agents can continue working.

Installation and running: requires the .NET 10 SDK (the SDK includes the runtime, so you only need to install the SDK) and Ollama. Having .NET 8 will not block you either: MandoCode provides both .NET 8 and .NET 10 builds, and dotnet tool install -g MandoCode will automatically choose the appropriate version for the machine; it is officially stated that this fallback mechanism will be removed after .NET 8 support ends in November 2026. You can run mandocode --doctor to see the runtime currently in use. The first run launches a setup wizard: it detects Ollama, prompts you to start it, guides cloud login (if you need a stronger model), and automatically pulls a reasonable default model; at any time you can use /setup Rerun. You can run /update within MandoCode to check for the latest stable version, and manual updates are also supported.

Main capabilities and usage:

1. Multiple agents and workspaces: Each task can be assigned an independent agent, with up to four independent agents per workspace, each having its own model, conversation, and approval settings; use Alt+N to open agents, Alt+Left/Right to switch agents, and Alt+Shift+N to create a workspace.
2. Portable context: Use /context-snap-create to save decisions, findings, and next steps from a conversation; use /context-snap-import (Alt+C) to search, preview, and import snapshots into another agent; snapshots are persisted in ~/.mandocode/snapshots.json and shared between CLI agents and workspaces.
3. Resume previous work: Use /history (Alt+H) to find saved agents by name, project, model, or conversation text, preview the conversation, and then restore their name, model settings, transcript, and conversation context.
4. Agent collaboration: Type @ at the prompt and then press Tab to switch from Files to Agents, allowing you to ask other agents about findings, request reviews, or hand off independent work; agent interactions have a separate collapsible area showing participants and real-time status, which can be expanded to view requests, messages, task comments, tool calls, and timestamped Markdown replies. Subtasks have readable IDs like job1, inbox messages are like message1, and tasks report outcomes such as Completed, Declined, Cancelled, or Failed, and return after being accepted rather than waiting for the result.
5. Project tools: Use Alt+E to open the file browser to browse project folders and files; selecting a file inserts an @ reference at the prompt cursor; use Alt+D to change the current agent's project directory, and Alt+G to view Git changes; each agent keeps its own project directory, allowing work across projects within the same workspace.
6. Extensions and integrations: /mcp (Alt+M) opens the MCP manager, where you can add or edit servers, test connections, view available tools, and enable or disable integrations; /skills (Alt+K) allows installing skills from Git, ZIP files, or local folders, and you can also write, generate, and refine your own instructions and then save them; shared integrations are available across agents, while each agent keeps its own conversation and settings.
7. Model management: Initial setup and /setup uses a guided panel (connect → select model → download → verify → ready), with the option to choose a recommended cloud starter model (such as glm-5.3-flash:cloud or deepseek-v4.1-flash:cloud), a local starter model with an approximate download size, an already installed model, or the full model browser; the cloud requires Ollama login and an internet connection, and the official notice states applicable account limits and pricing; local model memory usage depends on the model and context and exceeds the download size. The download target is the configured Ollama server; remote cloud authentication must be configured on that server. Setup verifies with a real model response; if verification fails, you can retry, switch models, or complete it later; existing model preferences are retained until verification succeeds. /ollama-pull lets you browse the Ollama library at any time, select model tags, and download with progress; the search box can filter model names, it supports Popular, Name A–Z, and Newest sorting, and toggling Cloud shows only models marked as cloud-hosted; /model is used to switch to an already downloaded model.
8. Terminal interaction: the conversation scrolls above a persistent multiline prompt, with a searchable selector and collapsible tool output; typing / shows an autocomplete dropdown and searches commands, typing ! runs shell commands; /keybindings shows the full shortcut reference (including menu navigation and editor controls).

Use cases: suitable for development work in the terminal such as code refactoring, diff proposals, codebase Q&A, cross-file search and planning, context handoff, parallel multi-agent tasks, MCP and skill extensions; it also provides the MandoCode Desktop native Windows interface (same engine, WinUI 3 on top, with an integrated shell, a Git-aware file browser, and 16 built-in themes), suited to users who prefer a native app.

Limitations and notes: requires the .NET SDK; local tools and source builds should use the original installation method; updates ask first before closing all workspaces and saving agent history (persistence must be enabled), the terminal prints the update log path, and mandocode must be restarted after completion; cloud models require Ollama login and an internet connection and are subject to account limits and pricing; local model memory usage depends on the model and context; shortcuts depend on whether the terminal forwards key combinations; agent handoff only communicates responsibility and file scope to the model and does not lock files; read-only review allows built-in inspection tools, but shell commands and unknown MCP tools are unavailable; a busy agent can receive messages and updates, but will only consume them on the next turn and will not inject them into the current active request; cancellation applies only to that interaction, and ownership is checked by the assigned agent. Other facts not provided here are not yet available.