features

eVi feature reference

A catalog of everything eVi does — what each feature is, how to use it from the CLI / REPL / Web, and where to configure it. Setup-heavy items link to Configuration; surface coverage is in Surface parity — CLI ↔ Web ↔ Desktop; the Claude-Code comparison is in eVi vs Claude Code — feature comparison.

Conventions: CLI = evi <cmd>; REPL = a /slash command inside evi chat; Web = the browser/desktop UI. Everything is local-first and opt-in unless noted.

---

Core chat

Interactive chat — evi chat (REPL) or the Web chat view. Streams tokens, runs tools, shows a working-status bar (spinner + elapsed timer + live token count) while a turn runs.

Session modes — Chat / Cowork / Code presets gate the tool set. REPL: pick at launch; Web: the mode switcher; per-turn headless: evi run -m code.

Headless — evi run "prompt" [-m mode] [-y] [--format json] [--schema f]. One-shot for scripts/CI. Example: git diff | evi run "review for bugs" -y.

Batch — evi batch prompts.jsonl -o out.jsonl -j 4. Runs many prompts (each its own agent), optional parallelism. Input is .jsonl/.json objects ({prompt, id?, mode?, schema?}) or one prompt per line.

Variants — evi variants "prompt" -n 3 — N independent takes on one prompt.

Tools

20+ built-in tools, each gated by a [tools] toggle in config.toml (Web: Settings → Tools): fs, code (run_python, sandboxable), shell, web (search/fetch), image (ComfyUI), vision, memory, pdf, sqlite, index (semantic search), git, ocr, calendar, computer, voice, mcp, subagent, federation. REPL: /tools lists active; /notools, /forcetool <name> per turn.

Sandboxed code — [tools] sandbox = true runs run_python under bwrap (Linux) / sandbox-exec (macOS): read-only FS + no network. Falls back gracefully.

Permissions & guardrails

Permission modes — [auto] mode = ask|accept_edits|plan|yolo + an auto_approve category list + first-match rules (allow|deny <tool> [arg]). REPL /auto, /plan. Trusted dirs/domains auto-approve fs/web under given paths/hosts. MCP allowlist (tools.mcp_allow) gates which servers load.

Content guardrails — ~/.evi/guardrails.toml, three layers: regex (block/redact), [[judge]] (the LLM classifies vs a policy), [[classifier]] (a local HF model, [moderation] extra). evi guardrails list|test. See Configuration.

Memory & context

Memory — the memory tool persists notes to ~/.evi/memory/; supports tags. Auto-recalled into context.

Context management — automatic compaction ([llm] compact_*). REPL /context (/ctx) shows a per-bucket breakdown (system/you/assistant/tools); Web: click the usage chip. /compact forces it.

Predicted outputs — /predict <text|file> speculative-decoding hint.

Sessions

Per-day JSONL transcripts (tools.transcripts). evi sessions list|show|resume| fork|continue|handoff; REPL /recent. Cross-device handoff: evi sessions handoff → resume on another machine after evi sync (or open /?session=<id>). Checkpoints/rewind: evi rewind (Web: rewind dialog) undoes file writes.

Skills, commands, styles

Slash commands — built-in /cmd controls plus ~/.evi/commands/<name>.md → /name templates (guide). Skills — ~/.evi/skills/<name>/SKILL.md instruction packets the model loads on demand via invoke_skill (guide). Output styles — evi style / [llm] output_style layer a persona onto the system prompt. Keybindings — ~/.evi/keybindings.toml maps a key to a REPL slash command.

Hooks

~/.evi/hooks.toml — run a command (argv) or POST a url around events. Tool events (before_tool_call/after_tool_call, glob match, veto_on_nonzero) and lifecycle events (user_prompt_submit/before_compact/stop). See Configuration.

Plugins

evi plugin add <dir|git-url> installs a bundle under ~/.evi/plugins/<name>/ that can carry commands/, skills/, hooks.toml, mcp.json, and agents.toml (subagent profiles), all auto-discovered. Marketplace: evi plugin search <q> / evi plugin install <name> resolve from ~/.evi/marketplace.json + [plugins] index_urls; evi plugin index init|add.

Agents & orchestration

Subagents — the delegate tool runs a scoped sub-agent by profile (built-in explore/plan + plugin/user agents.toml); evi agents lists them, evi agents new scaffolds a custom one. Agent teams — a shared, claimable task list (evi team new/run): a lead decomposes a goal into dependency-ordered tasks and teammates drain it in parallel (evi/teams.py). Parallel research — parallel_research tool fans out read-only explorers. Workflows — ~/.evi/workflows/<name>.toml, multi-step with parallel blocks + {step}/{var} interpolation; evi workflow new|run. Dispatch (Web 🗂) — lists live sessions + runnable workflows. Federation — delegate a task to a trusted peer eVi (~/.evi/peers.json, delegate_peer tool, [federation] serve to answer). Code review — evi review --multi multi-agent review. Multi-model routing — evi route picks a model per turn. Ultracode — evi ultracode "<task>" / /ultra / /effort ultracode runs one hard task through an exhaustive decompose → fan-out solvers → adversarial verify → synthesize pipeline (guide).

MCP

Client: ~/.evi/mcp.json servers, tools.mcp on; evi mcp list-tools. Server: evi mcp serve exposes eVi's tools/memory/commands over stdio+HTTP.

Automation

Recipes — evi recipe saved multi-turn flows. Routines — webhook → recipe (evi routine, POST /api/routine/{token}). Scheduled tasks — evi schedule add --cron … (--prompt | --eval <suite>); runs on cron, incl. scheduled evals for drift watch. Channels — push an external alert into a live web session (POST /api/session/{id}/channel).

Structured output & evals

Structured outputs — JSON-Schema-constrained replies: REPL /schema <file>, evi run --schema, Web /api/chat output_schema. Evals — evi eval new|run runs prompt→assertion suites (contains/regex/equals + LLM-judge rubric) with a pass-rate; exits non-zero to gate CI.

Voice & vision

TTS — [voice] engine = system|coqui|f5|piper; evi voice speak|engines; Web Settings → Voice; REPL /speak auto-speaks replies. STT — evi voice listen (faster-whisper, [stt] extra); always-on AutoListener. Vision — attach images (/image, Web 📎) to VLM models. Image gen — ComfyUI via the image tool.

Web & desktop

Settings — full screen over config.toml (9 sections). Multi-user — [web] multi_user + ~/.evi/users.json; each user gets an isolated workspace (sessions/transcripts/memory under ~/.evi/users/<name>/). Deep links — evi://session/<id> / evi://workflow/<name>; evi link. Desktop — native menus, tray, silent auto-updater (with in-app progress), first-run wizard. Status line — customizable REPL status ([statusline]).

Observability

OpenTelemetry — opt-in traces/metrics around tool calls ([telemetry] traces/metrics + otlp_endpoint, [otel] extra). Local stats — evi stats aggregates sessions/tools/tokens from transcripts. Crash reporting — opt-in Sentry ([telemetry] crash_reports + dsn, [telemetry] extra).

SDK (programmatic)

Agent SDK — evi.sdk is the curated public library surface: build_agent() (the convenience constructor the CLI itself delegates to), the streaming Agent

• all event types, @tool, subagents, structured output, run_headless, sessions/checkpoints, MCP, ultracode/workflows, and telemetry — all under one stable namespace (py.typed). Guide: Agent SDK (evi.sdk); runnable scripts: examples/python/.

Machine ops

Sync — evi sync push|pull git-syncs portable ~/.evi state. Backup — evi backup. Profiles — evi profile per-machine config overlays. Worktrees — evi worktree for parallel work. Doctor — evi doctor (Web: Help → Diagnostics). Fine-tune export — evi finetune export → JSONL training set.

Packaging & release

PyPI (v* tag, Trusted Publishing + sigstore signing), Docker image (docker.yml → GHCR), desktop installers (desktop-v*, 3-OS), a reusable evi-run GitHub Action, and CodeQL + gitleaks security scanning. See Releasing. To build locally in one command (sidecar + Tauri), run scripts/build-desktop.{ps1,sh} — see Self-build — developing and building eVi with eVi for developing/building eVi with eVi (the EVI.md bootstrap).

---

Generated from docs/features.md — edit there, not here.