ClydeCLI
Status
What works, what doesn't. Straight answer.
Core Systems
| System | Status | Description |
|---|---|---|
| CLI Entry | ✅ | clyde, login, logout, config, --model, --list-models, -c, --resume, -p (headless: answer on stdout, JSON output, exit status) |
| Interactive REPL | ✅ | Rich output, history, tab completion, multiline, streaming |
| Multi-Provider | ✅ | 13 providers (incl. local Ollama and oMLX), stdlib HTTP, live model lists, provider:model switching |
| Agent Loop | ✅ | Tool-calling loop with retries, reasoning control, history repair |
| Skill System | ✅ | SKILL.md slash-command skills with args + tool limits |
| Context Building | ✅ | Workspace snapshot, git status, a README excerpt, entry points, the code map and memory files go into the prompt: ~/.clyde/CLYDE.md (user), CLYDE.md (project, shared) and CLYDE.local.md (project, personal, keep it gitignored). Files written for other agents are read too: CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, .github/copilot-instructions.md (one per folder, first found wins) |
| Permissions | ✅ | Bash asks for any command that is not read-only (dangerous patterns are refused), Write/Edit/NotebookEdit ask before every change in hold mode, Config asks before a change, WebFetch asks per domain and WebSearch asks before each search (showing the query). Answer "don't ask again" or add permissions.allow / deny rules (Claude Code syntax) to ~/.clyde/settings.json |
| Shell asks | ✅ | Commands that are not read-only ask; so does any command that names a folder outside the project, even in all-in mode. No OS sandbox |
| Sessions | ✅ | Auto-saved after each turn; /resume picker per workspace, clyde -c / clyde --resume [id] |
| Checkpoints | ✅ | /rewind undoes the model's file edits and/or the conversation back to before any of your messages; saved per session, so they survive /resume |
| Cost Tracking | ✅ | /cost shows input, output and cache tokens per model with an estimated $ total from catalog prices |
| Hooks | ✅ | PreToolUse / PostToolUse shell commands from ~/.clyde/settings.json or .toml; also reads Gemini CLI, Cursor and Copilot CLI hook tables |
| Plugins | ✅ | clyde plugin install bundles of tools, skills, hooks and MCP servers (Claude Code layout) into ~/.clyde/plugins/, enabled only after a yes |
| Self-checks | ✅ | After a Python edit, ruff (and mypy if configured) run on the file and new problems go back to the model; /check runs ruff, mypy and pytest |
| Tracing | ✅ | Every session writes ~/.clyde/traces/<session>.jsonl (model and tool calls, timings, tokens, secrets redacted); /debug shows the last turn, clyde --debug streams it live |
| Compaction | ✅ | /compact on demand; runs automatically once history reaches 80% of the context window |
| cardShuffle | ✅ | A model that deals each turn to an /eval-ranked real model (high-roller, house, free, small), with fallback on errors, max tool turns and stuck loops |
| Local models | ✅ | /models local finds ollama.com, Hugging Face GGUF and MLX models that fit this machine, rated relax / balance / hard, confirmed before download through Ollama (ollama.com, GGUF) or oMLX (MLX) |
| whichllm | ✅ | Bundled benchmark ranking of local models in /models local: the best models that fit this machine, with quality, speed range and license in the download confirmation |
| SkillSpector | ✅ | Bundled scanner for skills, plugins and MCP servers from other agents: static always, LLM review when a usable model is connected; DO_NOT_INSTALL items held back until /skills allow |
| Laya | ✅ | Optional local decision model (clyde-cli[laya]): hands a stuck cardShuffle turn to the next card; scores every turn's difficulty and steers house once /laya's evidence shows it separates easy turns from hard ones |
| IDE integration | ✅ | clyde --acp speaks the Agent Client Protocol, so editors that support it (Zed, JetBrains through an ACP adapter) run Clyde as their agent: streamed replies, tool calls and permission asks in the editor |
| Pasting | ✅ | Ctrl+V pastes a copied image, a copied image path becomes the image, long pastes fold to [Pasted text #N +X lines] |
Tools
| Category | Tools | Status |
|---|---|---|
| Files | Read, Write, Edit, NotebookEdit, Glob, Grep | ✅ Working; Edit also lands a unique match that is off only by trailing whitespace or indentation, and says so in its result, in the one-line summary and in the trace |
| Data | Data | ✅ Profile CSV, TSV, Parquet, JSON and SQLite files (types, nulls, distinct counts, min/max, samples) and run read-only SQL across them with DuckDB, sandboxed to the workspace |
| System | Bash | ✅ Working; asks before commands that are not read-only or that name a folder outside the project |
| Web | WebFetch, WebSearch | ✅ Working |
| Interaction | AskUserQuestion, SendUserMessage | ✅ Working |
| Tasks | TodoWrite, TaskCreate/Get/List/Update/Output/Stop | ✅ Working |
| Planning & config | EnterPlanMode, ExitPlanMode, Config, Skill, ToolSearch, Sleep | ✅ Working |
| Agent | Agent | ✅ Sub-agents on a fresh conversation: general-purpose or custom types (.clyde/agents/*.md, Claude Code's format), in the foreground or background (TaskOutput collects, TaskStop cancels) |
| Scheduling | CronCreate/List/Delete | ✅ Session-scoped; due jobs run as a turn while the REPL is idle at the prompt |
| Team | TeamCreate/Delete | ✅ Runs several sub-agents in parallel, each on its own prompt and type; TaskOutput collects their answers |
| Worktree | EnterWorktree/ExitWorktree | ✅ Creates a git worktree on a new branch; exit keeps or removes it |
| MCP | MCP, ListMcpResources, ReadMcpResource, mcp__<server>__<tool> | ✅ Stdio, Streamable HTTP and HTTP+SSE servers from ~/.clyde/settings.json; header auth or OAuth sign-in (/mcp login) |
| Code map | Map | ✅ query / path / explain / affected / god_nodes over a map of the repo, refreshed when ClydeCLI starts |
| LSP | LSP | ✅ Definition, references, hover, symbols and call hierarchy via a language server on PATH |
Roadmap
- ✅ Phase 0: Installable, runnable CLI
- ✅ Phase 1: Core agent experience (REPL, sessions, slash commands)
- ✅ Phase 2: Real tool-calling loop, multi-provider
- ✅ Phase 3: Context, permissions, recovery (context building, saved permission rules,
/resume,/doctor, compaction, hooks) - ✅ Phase 4: MCP client (stdio, Streamable HTTP, HTTP+SSE), plugins, custom tools/skills/hooks, tracing and
/debug - ✅ Phase 5: Python-native differentiators: the Code Map, check-after-edit with ruff/mypy/pytest/uv, notebook tools, and the Data tool for data and ETL work
- ✅ Phase 6: Model play: cardShuffle routing across
/eval-ranked models, local models that fit the machine (ollama.com, Hugging Face GGUF and MLX), Laya bundled for stuck-loop detection, and SkillSpector bundled to scan skills, plugins and MCP servers - ✅ Laya's difficulty in
cardShuffle:house: promotes itself from shadow mode once/laya's evidence shows it separates easy turns from hard ones - ✅ IDE integration through the Agent Client Protocol (
clyde --acp) - 🔲 Messaging from your phone (Signal, then maybe iMessage): proposed, not built. Design, security model and the upkeep each needs: messaging.md
See FEATURE_LIST.md for detailed feature status and PR guidelines.