A local-first coding workspace for running local models and subscription coding agents side by side.
Get started · Features · Architecture · Contributing
Klide runs Ollama, MLX, Claude Code, Codex, OpenCode, and Oh My Pi (OMP) without hiding their native capabilities. You can inspect tool calls, follow terminal output, compare model cost and latency, and review changes from one workspace.
Klide keeps three questions in view:
- What is running? Mission Control tracks Klide runs and delegate command-line interface (CLI) sessions
- What needs attention? Waiting, blocked, failed, and reviewable work rises to the top
- What changed? Each run keeps its files, branch, transcript, validation state, cost, and memory evidence
Coding agents work well in terminals, but their sessions, diffs, and evidence spread across separate tools. Klide keeps the real terminal experience and adds a shared operations layer around it.
| Run real agents | Use Claude Code, Codex, OpenCode, OMP, or a custom CLI inside persistent pseudoterminal (PTY) sessions |
| Keep work visible | See working, waiting, blocked, and completed runs across the workspace |
| Review before trust | Approve commands, inspect proposed edits, comment on diffs, and check validation evidence |
| Choose where inference runs | Start with Ollama or MLX, then opt into hosted providers or subscription CLIs |
| Resume with context | Reopen transcripts, hand work to another agent, and save reviewed project memory |
- Start a Klide run or open a delegate CLI
- Follow its status, tool calls, terminal output, and changed files
- Review edits and commands before they cross the workspace trust boundary
- Validate the result, send feedback, resume later, or hand the work to another agent
Klide has three capability modes:
| Mode | Agent access |
|---|---|
Chat |
Answers without workspace tools |
Plan |
Reads the workspace and proposes an approach |
Goal |
Uses tools with diff review and permission-gated commands |
| Area | Included |
|---|---|
| Agent operations | Mission Control, shared run lifecycle, attention queue, transcripts, session resume, cross-agent handoff, sub-agent visibility |
| Review and evidence | Diff comments sent to agents, command approval, checkpoints, validation status, files touched, tokens, cost, and stop reasons |
| Parallel work | Git worktrees, worktree setup recipes, branch comparison, and merge controls |
| Editor and shell | Monaco editor, file explorer, tabs, search, command palette, Git review, commit graph, and persistent PTY terminals |
| Models | Ollama, MLX, Anthropic, OpenAI, Mistral, xAI, OpenRouter, and OpenAI-compatible endpoints |
| Project context | AGENTS.md, CLAUDE.md, file mentions, skills, dynamic tools, and reviewed project memory |
| Local security | Workspace-rooted file access, operating-system keychain storage, project command allowlists, and network permissions |
Klide currently targets macOS. Apple Silicon is the primary development platform.
Clone the repository and start the Tauri development build:
git clone https://github.com/pierreprudh/KLIDE.git
cd KLIDE
npm install
npm run tauri devThe first Rust build takes about three to five minutes. Later builds reuse the compiled dependencies.
Once Klide opens:
- Open a workspace folder
- Open the AI panel and select a provider
- Press
Tabto switch betweenChat,Plan, andGoal - Open Mission Control to follow active and completed work
⌘P jumps to a file, ⌘⇧P opens the command palette, and ⌘/ shows the full shortcut cheatsheet.
pierreprudh/klide-8b is a Low-Rank Adaptation (LoRA) fine-tune of LFM2.5-8B-A1B. It is trained for Klide's tool and edit contract.
ollama pull pierreprudh/klide-8bThe model appears in Klide's Ollama picker after the download finishes.
Klide separates the interface from the durable execution layer:
| Layer | Responsibility |
|---|---|
| React and TypeScript | Editor, panels, review surfaces, layouts, and run projections |
| Tauri and Rust | Agent harness, provider streaming, tools, permissions, PTYs, Git, keychain, and filesystem access |
| JSON Lines (JSONL) transcripts | Append-only run events used by Mission Control, validation, memory, and replay |
The Rust harness checks capabilities when it advertises tools and again before execution. Writes pause for diff review. Commands and network access pause for permission.
Read the Harness contract for the trust model and Harness schema for the tool interface.
Klide is at version 0.5 and remains under active development. Source builds are available now; signed application bundles are not yet published.
Current priorities:
- Make worktree-per-run isolation the default parallel-agent flow
- Deepen the review queue and agent feedback loop
- Publish signed macOS builds, then validate Windows and Linux
See the changelog for shipped milestones.
Run the relevant checks before submitting a change:
npm test
npm run build
cargo test --manifest-path src-tauri/Cargo.tomlThe frontend lives in src/. The Rust backend lives in src-tauri/.
Issues and pull requests are welcome. Keep changes focused, preserve the workspace trust boundary, and include validation for changed behavior.
Klide is built with Tauri, Monaco, and xterm.js. Its product influences include Sinew, Ara, Cursor, Cline, Linear, and the open coding-agent ecosystem.
Klide is available under the MIT License.