claude code comparison

eVi vs Claude Code — feature comparison

How eVi's capabilities line up against the full Claude Code docs surface (reviewed 2026-06-16, eVi 0.33.0). eVi is local-first, single-user, privacy-first, so a number of Claude Code's cloud/enterprise features are intentionally out of scope rather than "missing".

Legend: ✅ have · ⚠️ partial · ❌ gap (buildable, on the roadmap) · 🚫 not planned (philosophy mismatch / separate big track).

Core

Claude Code	eVi equivalent	Status
Agentic loop + built-in tools	evi/llm/agent.py; 20+ tools	✅
Common workflows (explore/fix/refactor/test)	chat, evi review, recipes	✅
Interactive mode + keyboard shortcuts	prompt_toolkit REPL (tab-complete, history, keybindings.toml)	✅
Goals	/goal, agent.goal	✅
Scheduled tasks	scheduler (APScheduler), evi scheduled	✅
Voice dictation	STT + voice loop	✅

Sessions & conversation

Claude Code	eVi equivalent	Status
Name / resume / branch / switch	evi sessions (list/show/resume/continue/fork/title); edit/branch/reroll	✅
Checkpointing (rewind edits + convo)	file rewind (Ph 64) + convo edit/branch/reroll	✅
Continue / resume / fork	evi sessions continue / fork (Ph 71)	✅
Remote control (drive from a phone)	remote-backend mode + multi-machine	⚠️ no cross-device session handoff
Deep links (open from a URL)	—	🚫 niche

Permissions & safety

Claude Code	eVi equivalent	Status
Permission modes (ask/accept-edits/plan/yolo)	auto.mode (Ph 66)	✅
Fine-grained allow/deny rules	auto.rules (Ph 66)	✅
Auto-mode config (trusted dirs / domains)	auto.trusted_dirs + trusted_domains (Ph 77)	✅
Sandboxed bash/code	[tools] sandbox (Ph 67)	✅
Security / vuln review	guardrails, security.yml (pip/cargo audit), review --multi	✅
Content moderation / safety filter	guardrails.toml — regex (block/redact) + LLM-judge + offline ML classifier + dedicated guard model (Llama Guard / ShieldGemma) on input/output	✅ (no in-harness equivalent in Claude Code)

Extension & customization

Claude Code	eVi equivalent	Status
CLAUDE.md / memory	EVI.md/AGENTS.md project context + memory (+ tags)	✅
Nested CLAUDE.md (large codebases)	merges EVI.md/AGENTS.md up the tree (Ph 76)	✅
Skills	SkillStore	✅
Subagents	delegate_* + parallel_research (Ph 61)	✅
Output styles	[llm] output_style (Ph 69)	✅
Hooks (incl. HTTP hooks)	hooks.toml (command/url; tool + lifecycle events: user_prompt_submit/before_compact/stop/session_start/session_end; conditional arg_match)	✅
Status line	[statusline] (Ph 72)	✅
Keybindings	keybindings.toml + evi keybindings (Ph 82)	✅
Skills tool-scoping (allowed/disallowed-tools)	SKILL.md allowed-tools/disallowed-tools → evi/skillscope.py scopes the toolset while a skill is active	✅
Custom slash commands	~/.evi/commands frontmatter/$ARGS/@file/namespacing (Ph 62)	✅
Nested skills (subfolders)	recursive SKILL.md scan	✅
/add-dir (extra working dirs)	/add-dir → session trusted_dirs	✅
!cmd shell passthrough	!cmd in the REPL (output folded into context)	✅
AskUserQuestion (clarifying Qs)	ask_user tool (interactive-only; graceful no-op in web/headless)	✅

Plugins & MCP

Claude Code	eVi equivalent	Status
Plugins (bundle commands/skills/agents/hooks)	evi plugin add/init/list/remove; bundles commands+skills+hooks+MCP+subagent profiles; installs from dir / git / .zip / URL; bin/ on PATH	✅
Plugin scaffolding	evi plugin init <name> (starter command + skill)	✅
Plugin marketplaces / discovery	evi plugin search/install + [plugins] index_urls (curated index)	✅
MCP client (connect to servers)	MCPManager, mcp.json	✅
MCP server (publish eVi's tools)	evi mcp serve (stdio + HTTP)	✅
Managed MCP (allowlists + output cap)	per-tool allow on publish + tools.mcp_allow (Ph 78) + tools.mcp_max_output_chars	✅

Platforms

Claude Code	eVi equivalent	Status
CLI	Typer CLI	✅
Desktop app	Tauri desktop (menus/tray/updater)	✅
Web	FastAPI + SSE web UI	✅
Headless / print mode	evi run (Ph 65)	✅
VS Code / JetBrains extensions	—	🚫 separate track
Mobile / web cloud platform	—	🚫 local-first
Computer use	computer tool	✅
Chrome / browser automation	—	🚫 deferred for MCP browser servers
Slack integration	—	❌ integrations backlog

Review · planning · orchestration

Claude Code	eVi equivalent	Status
Code review (multi-agent)	evi review --multi (Ph 70)	✅
Run agents in parallel	parallel_research (Ph 61)	✅
Worktrees (isolate parallel sessions)	git worktrees + [worktree] base_ref default	✅
Dynamic workflows (script many subagents)	recipes + workflows + ultracode	✅ TOML DAG + fixed pipeline
Agent teams (shared task list)	evi team — lead decomposes, teammates claim+drain (evi/teams.py)	✅
Agent view (live dispatch dashboard)	Dispatch panel + GET /api/dispatch/stream (SSE, busy dots)	✅
Ultrareview (CI gate)	evi review --multi --exit-code / --json + /ultrareview	✅ local
Ultraplan (cloud)	local plan mode	🚫 cloud

Automation & triggers

Claude Code	eVi equivalent	Status
Routines (schedule / webhook trigger)	scheduler + evi routine webhook→recipe (Ph 73)	✅
Channels (push alerts into a session)	POST /api/session/{id}/channel — note by default, or run:true drives a live agent turn	✅
GitHub Actions / GitLab CI	packaged evi-run composite action + ready examples/github/pr-review.yml	✅

Config · models · cost

Claude Code	eVi equivalent	Status
Settings (global + project)	config + profiles + .evi.toml (Ph 74) + settings UI	✅
Env vars / .claude dir	EVI_* env + ~/.evi	✅
Model config / aliases / fast mode	model picker, routing, fast_mode	✅
Fallback model	[llm] fallback_models (retry the turn down the chain)	✅
Extended thinking on/off	reasoning_effort off/low/medium/high/max (/effort)	✅
Transcript retention (cleanupPeriodDays)	tools.cleanup_period_days + evi sessions purge	✅
Prompt caching	cache_prompt	✅
Context window display	usage chip + status line	⚠️ no interactive sim
Cloud backends (Bedrock/Vertex/Foundry)	openai_compat (covers gateways/proxies)	🚫 by design
Cost management / analytics dashboards	token usage shown	🚫 local = free

SDK & programmatic

Claude Code	eVi equivalent	Status
Headless mode	evi run --format json (Ph 65)	✅
Custom tools	@tool decorator	✅
File checkpointing	evi rewind (Ph 64)	✅
Structured outputs	/json + JSON-Schema (/schema, evi run --schema)	✅
Batch API	evi batch <file> → JSONL (parallel)	✅
Evals	evi eval (assertions + LLM-as-judge; --eval on a schedule)	✅
Usage analytics	evi stats (local; sessions/tools/by-category/busy days)	✅ local-only
Transcript search (Ctrl+R / resume)	evi sessions search <query> (snippets)	✅
Responses API built-in tools	[llm] responses_tools (web_search/code_interpreter/…)	✅ opt-in
Multi-user / teams	[web] multi_user + users.json (per-user tokens + isolated sessions/transcripts/memory)	✅ opt-in
Federation (agent↔agent across machines)	evi peer / delegate_peer / /api/federate	✅ eVi-unique
Permissions SDK	permission policy (Ph 66)	✅
Public Agent SDK (library)	evi.sdk — curated re-export + build_agent() + examples (Agent SDK (evi.sdk))	✅
Session storage (S3/Redis)	local JSONL transcripts	🚫 local-first
OpenTelemetry / monitoring	opt-in crash reporting (Sentry-compatible)	⚠️ no metrics/traces
Tool search (1000s of tools)	[tools] tool_search defers the long tail behind a search_tools meta-tool	✅ opt-in

Org / enterprise / compliance

Claude Code	eVi equivalent	Status
Admin setup, server-managed settings, team onboarding	—	🚫 single-user
Authentication (accounts/SSO)	web auth token; local backends need none	🚫 local
ZDR / legal / data-usage	private by design (everything local)	🚫 N/A

Summary — specialty SLMs + opencode/Cursor gleanings (0.34.0)

Beyond Claude Code, eVi pulled in the genuinely-local wins from opencode and Cursor and added small specialty models:

• Specialty SLMs — [models] ocr/vision/stt/tts/guard/diarize/doc_layout registry; describe_image + OCR-VLM routing (Moondream2 / Qwen2.5-VL / GLM-OCR), Kokoro TTS, faster-whisper turbo. A small model per task, no main-model swap. (0.37.0) added a safety-guard layer (Llama Guard / ShieldGemma → [[guard]] rule), speaker diarization (evi voice diarize, pyannote), and document layout/OCR (ocr_image engine=doc, Docling). Capability chips now cover 🔧 tools, 🛡 guard, ◆ embeddings/reranker (plus the existing 👁 vision / 🧠 thinking / ⌨ infill / 🎤 audio).
• Working folder — per-session cwd (/cd, --cwd, web 📁 chip).
• opencode core — real shell tool, apply_patch (multi-hunk), format-on-edit + check_file diagnostics (LSP-lite), persistent /plan toggle, evi init (AGENTS.md, already discovered up-tree).
• Cursor gleanings (local-first only) — local FIM completion engine (evi complete / /api/complete) so eVi is a local Tab/Copilot backend; Bugbot-style review (.evi/BUGBOT.md + evi review-remember + severity). (evi edit already existed.) Cursor's cloud bits (Cloud Agents, remote PR Bugbot, Design Mode, Slack, enterprise) remain out by the local-first rule.
• Deferred (need a separate client / deeper work): a VS Code/LSP extension to render the FIM completions as ghost-text; a full language-server integration (eVi ships the lighter check_file/format-on-edit instead).

Summary — S/M parity batch (0.33.0)

✅ Closed in 0.33.0 (the "close every buildable small/medium gap" pass):

• Model fallback chain — [llm] fallback_models retries the turn down the chain on a setup failure.
• Extended thinking off — reasoning_effort = "off" (plus /effort off); centralized in reasoning.py.
• Transcript retention — tools.cleanup_period_days (auto-prune on startup) + evi sessions purge.
• Transcript search — evi sessions search <query> with snippets.
• MCP output cap — tools.mcp_max_output_chars truncates chatty tool results.
• Conditional hooks — arg_match gates a hook on tool arguments, not just the tool name.
• Session lifecycle hooks — session_start / session_end.
• CI-gating review — evi review --multi --exit-code / --json + /ultrareview.
• Plugins — evi plugin init scaffold; install from .zip/URL; bin/ on PATH; recursive (nested) skill discovery; /reload-skills.
• /add-dir — trust an extra directory for the session.
• !cmd — REPL shell passthrough (output folded into context).
• ask_user tool — AskUserQuestion parity (interactive-only, graceful no-op elsewhere).
• worktree.base_ref — default fork point for evi worktree create.
• Usage by category — evi stats attributes tool calls per category.

✅ Already shipped (0.31.0 → 0.32.0): the full Phase 75–94 roadmap (plugin skills/hooks/MCP/subagent-profiles, nested project context, trusted dirs/domains, MCP allowlist, keybindings, channels, packaged CI action, cross-device handoff, context-window breakdown, OpenTelemetry, fine-tune export, voice engines, CodeQL + gitleaks, Docker→GHCR, sigstore signing), plus evi:// deep links, the plugin marketplace index, and the public Agent SDK (evi.sdk; see Agent SDK (evi.sdk)).

⚠️ Deferred (need an architectural addition or are low-value), not philosophy gaps:

• Skill tool-scoping (allowed-tools/disallowed-tools in skill frontmatter) — needs a persistent "active-skill" mode in the agent loop (skills are currently one-shot instruction packets).
• Nested subagent spawning (subagents that spawn subagents) — deliberately gated as a runaway-cost guard.
• Custom REPL themes — terminal colour themes (output styles already cover response persona).

🚫 Not planned (by design): cloud/enterprise backends (Bedrock/Vertex/Foundry), IDE extensions, mobile, agentic browser, hosted session storage, cost/analytics dashboards, accounts/SSO/admin — eVi is local-first and single-user.

---

Generated from docs/claude-code-comparison.md — edit there, not here.