CodeSharp.Cli 0.9.14-alpha

CodeSharp

A C#/.NET 10 agent harness built on Microsoft's Agent Framework (MAF), with a specialism in .NET development. Runs locally with zero required infrastructure and scales additively — Aspire, Docker, MCP servers, a web UI, and multi-agent orchestration are all opt-in, never required for the basic local workflow.

Status: alpha (0.9.x), cross-platform. Actively developed; behavior, config shape, and CLI surface can still change between releases. See Platform support below.

Quick start

dnx CodeSharp.Cli --prerelease -y

dnx (bundled with the .NET 10 SDK) runs the tool without a separate install step. --prerelease is required today since no non--alpha version has shipped yet. First run walks you through a setup wizard (provider/model selection, tool toggles). If you already have a known provider key exported (e.g. NVIDIA_NIM_API_KEY, CLOUDFLARE_API_TOKEN), CodeSharp notices at startup and offers to set that provider up for you — never silently. See Pre-setting a key before first run.

To install it as a persistent global tool instead:

dotnet tool install -g CodeSharp.Cli --prerelease
codesharp

What it does

CodeSharp is an interactive, terminal-first coding agent (Spectre.Console-based TUI) with:

• Persistent named sessions — /new, /resume, /sessions, with full conversation history restored across restarts.
• Model provider flexibility — native Anthropic support (API key or claude setup-token subscription auth), OpenAI-compatible endpoints, Ollama for local/local-network models, and a provider registry (/provider add|list|edit|remove) with guided setup for several presets (including assisted key acquisition for providers like NVIDIA NIM).
• Model tiering — beyond the default model, dedicated tiers for an advisor (/advisor, consult without acting), escalation (/escalate, hand a task to a stronger model), a cross-provider reviewer (/review), and context compaction (/compact) — each independently configurable, so e.g. a fast local model can review the output of a hosted frontier model.
• In-place context compaction (/compact) — summarizes the middle of a long conversation while keeping the earliest and most recent turns verbatim, without starting a new session.
• Built-in tools — file read/write/edit, grep/glob search, shell execution (workspace-confined, permission-gated), git inspection and first-class commit/branch/PR tools, web search/fetch, background process management, notebook (.ipynb) cell editing, and a scripting ladder from one-off scripts up to versioned, sandboxed local modules.
• Multimodal input — read_image for image files, clipboard paste (/paste), and read_pdf for PDFs: text extraction by default, or render_pages to rasterize selected pages so a vision model can read a scanned or image-only document. Rendering is fully cross-platform (PDFium, no system dependency) and bounded by page/DPI/size caps; it is offered only to vision-capable models.
• MCP support — both directions: connect to external MCP servers (stdio and HTTP transports, including Docker-managed server lifecycle) and expose CodeSharp's own tools over MCP.
• Skills and sub-agents — Agent Skills (spec-compliant), named local sub-agents, and A2A (Agent2Agent) client/host support for cross-process agent orchestration, including sequential/concurrent workflows with human-in-the-loop approval.
• Memory — opt-in semantic recall over past conversation turns and durable notes, plus a scheduled background "dreaming" pass that consolidates history into higher-level notes.
• Feedback capture — built-in good/bad feedback capture with model metadata, exportable for fine-tuning (SFT/DPO formats).
• Permission model — a risk-classed tool-approval gateway (Safe/Moderate/Destructive), plan mode (mandatory review-before-action), and a --yolo escape hatch for trusted automation.
• Session continuity — /handoff + /resume (and auto-refresh) for carrying context across a fresh session when the current one gets long; scheduled wakeups (--task <id> + OS-level scheduling) for headless, non-interactive runs.
• Web UI (optional) — attach from a browser on the same device or LAN (QR-code pairing), alongside the terminal session.

Full configuration reference: see docs/configuration.md in the source repository.

Platform support

Each release ships packages for win-x64, linux-x64, osx-x64, and osx-arm64 (plus a portable package). CI builds and runs the automated test suite on all three OS families (windows-latest/ubuntu-latest/macos-latest) on every push/PR to main — the Linux leg runs every test class; the Windows and macOS legs exclude the two Docker-dependent classes (DockerScriptSandboxTests/DockerBackgroundProcessTests), a known, documented gap. The release workflow (tag push) re-runs build + test independently, but on windows-latest only and with the same Docker-class exclusion — it is not a three-OS gate. OS-specific code paths (shell invocation, credential storage, path handling) each have a real, dedicated implementation per platform — e.g. DPAPI on Windows, the freedesktop Secret Service (libsecret) on Linux, and Keychain on macOS for credential storage; pwsh on Windows and bash/sh elsewhere for shell execution.

Documentation

• docs/configuration.md — the exhaustive operator/end-user configuration and feature reference (permissions, providers/credentials, MCP, memory, web UI, plan mode, feedback export, and every other config key).
• docs/claude-code-command-differences.md — read this first if you are arriving from Claude Code: the five commands (/rewind, /cost, /init, /compact, /resume) whose names CodeSharp shares but whose behaviour differs.
• docs/architecture/decisions.md — the locked architectural decisions (ADRs): IChatClient as the sole model seam, MAF as the engine, MCP as the plugin model, shell-first tooling, model tiering, Aspire-optional core, and more.
• AGENTS.md — orientation for contributors and AI coding agents working in the codebase: the project map, conventions, build/test commands, and the spec-driven development workflow. It links focused topic docs under docs/agents/.

Contributing

Development follows a spec-driven cycle (brainstorm → spec → plan → test-first execution), with the design record committed under docs/. Start at AGENTS.md. The project currently operates directly on main (alpha, no live users yet) — see the development-workflow doc for details.

License

Not yet declared — no LICENSE file exists in the repository yet. Treat this package as all-rights-reserved until a license is published; if you need clarity for your use case, open an issue on the GitHub repository.

Links

• Source & issues: github.com/jamesburton/codesharp
• Package: nuget.org/packages/CodeSharp.Cli