Local-first AI agent OS for your machine. Daemon mode in sub-50ms, swarm runner across isolated git worktrees, sandboxed self-dev that only ships if tests pass, voice dictation, mermaid in the dashboard, and a one-shot importer that pulls your context out of Claude Code, Codex, or OpenCode. Tiered sandbox, signed plugins, multi-channel reach, Memory v2 — all built in. One product. One price. Yours forever.
Phantom v1.1.35 ships in two tiers. Free covers chat, plugins, memory, MCP, bench, doctor — all forever, no card needed. Pro (₹999 lifetime) adds daemon mode, swarm runner, sandboxed self-dev, and voice dictation. Pro features are unlocked free for the first 14 days so you can try them. Existing licence holders just /update; pre-gate users are grandfathered. Every future patch free, on both tiers.
/update.phantom serve opens a unix-socket daemon; phantom connect attaches in sub-50 ms. Closes the cold-start gap against Rust harnesses without a rewrite. macOS connect-before-accept race fixed with a warm-up retry ping.
phantom swarm "<goal>" --agents N fans N subagents into isolated git worktrees, runs them in parallel, collects diffs, and flags file-level conflicts before merge. Try the same task three ways and pick the winner.
phantom self-dev "<change>" applies an edit in a worktree, runs your full test suite there, and only swaps onto your branch (with --swap) if everything stays green. Iterate without ever breaking main.
phantom dictate records via sox/arecord/parecord and transcribes through Whisper. Stub backend ships for offline tests so CI doesn't need a microphone. On top of the existing voice loop with VAD-driven STT/TTS and barge-in cancellation.
phantom memory import {claude-code,codex,opencode} pulls transcripts from other agents straight into Phantom's episodic memory — switch tools without losing context. Pair with phantom mcp import which slurps ~/.claude/mcp.json and ~/.codex/mcp.json in one command. No more re-pasting server configs.
phantom bench prints cold-start time, daemon round-trip, RSS, turn-latency p50/p95, and scaling slope as JSON. Methodology baked in — anyone can verify the sub-50 ms claim on their own machine.
phantom config provider custom <name> --base-url --model --key-env registers vLLM, Ollama, LM Studio, or any OpenAI-shaped endpoint in one shot. No code edits, no JSON hand-editing.
Fenced ```mermaid blocks render live in the web dashboard via the official renderer with securityLevel: 'strict' and a MutationObserver for streamed inserts. Architecture diagrams arrive drawn, not as ASCII.
github-pr — gh CLI inspector for any PR. web-screenshot — Playwright PNG capture of any URL. code-review — pure-Python static lints over a unified diff. Demonstrates the Ed25519-signed plugin SDK end-to-end. Existing 5 (clock, weather, gh-search, code-search, todo) carry forward — eight first-party plugins total.
Tiered sandbox (bubblewrap → firejail → unshare → docker on POSIX, passthrough on Windows v1.0), Ed25519-signed plugin SDK, multi-channel adapters (WebChat / Telegram / Discord / Slack / Matrix), MCP both-ways (client + server), ACP child-agent runtime with topological waves, Skills (SKILL.md bundles), Memory v2 (SQLite + FTS5 + TF-IDF reranker), realtime voice with barge-in, typed Canvas UI, PWA + service worker, i18n in 5 languages, KeyPool round-robin rotation, OTel-shaped metrics. Release date: 2026-05-05. ₹999 lifetime — every v1.0.2 capability included, every future patch included.
Existing users — your licence and install path are preserved. Run /update in your REPL. Both phantom (canonical) and python run.py (alternate entry point) keep working.
Designed to execute tasks on your local machine with real OS access, permanent memory, and remote control from your phone.
A heavy Main Engine handles reasoning, coding, and execution. A fast Router Engine classifies your prompt and shapeshifts into the right specialist — File Organizer, DevOps Agent, Code Writer, or Security Auditor. Works with any OpenAI-compatible provider.
Real bash commands on your machine. 48 catastrophic patterns permanently blocked. Trust levels 1–4. Level 4 requires your PHC license key to unlock — and auto-downgrades to Trust 3 after 30 minutes idle.
Full two-way chat with PhantomCLI from your phone. All /commands work. Separate trust level from local CLI. Owner-only security gate.
Image generation (FAL.ai, DALL-E, Stability, Replicate), voice TTS (ElevenLabs, OpenAI, PlayHT), speech-to-text (Deepgram). Configure once, use via natural language.
One purchase activates up to 3 machines. Hardware fingerprinting + server-side enforcement. Manage devices via python run.py setup.
Boot sequence, HUD panels, Claude Code-style thinking spinner. Local web dashboard at localhost:8080. Nothing hosted externally — your data stays on your machine.
Every tool call — shell, file write, search, edit — is shown as a live checklist (◻ → ◼ → ✔ / ✖) in both the terminal and the web dashboard. For multi-step jobs the model plans upfront, so you see exactly where it is.
Hand it a dataset — it auto-installs pandas, scikit-learn, xgboost, plotly, transformers, whatever fits the task. Writes modular Python, trains the model, serves an interactive FastAPI / Flask / Streamlit dashboard and prints the URL.
PreToolUse, PostToolUse, UserPromptSubmit, Stop, SubagentStop, SessionStart/End, PreCompact, Notification. Shell hooks (stdin JSON) + programmatic Python hooks via @hook decorators. Non-zero exit blocks for binding events; fail-open on timeout.
Scoped allow/deny/ask: bash:git:*, write:~/projects/**, deny:bash:rm:*. Every decision hash-chained into ~/.phantom/audit.jsonl. Tamper-detectable with /audit verify.
Auto-loads CONTEXT.md from three scopes (user, project-root, local dir) into the system prompt on session start. Poll-based watcher picks up edits live — no restart needed.
Classifier picks cheap tier (tool-arg turns), mid tier (default), or expensive tier (reasoning cues) per turn. Cuts token cost 40–60% vs single-model. Per-model price sheet + daily spend alerts via the Notification hook.
Dedicated activity bar with 4 tree views (Sessions, Memories, Hooks, Tools). 5 code actions on any selection: explain, refactor, fix, add tests, document. Quick Pick palette, rich status bar with live cost tile, workspace context injection (active file + open files + diagnostics), transcript export to Markdown, inline diff preview. Stdio bridge to the CLI.
Model Context Protocol client with resources/list, resources/read, subscribe/unsubscribe, and auto-reconnect with exponential backoff. Every MCP tool call surfaces in the model's structured tool list.
Native Anthropic SSE (message_start / content_block_delta / message_stop) AND OpenAI-compatible chunks. UTF-8 chunk-boundary safety, partial-JSON repair for tool args truncated mid-stream. 60+ combined tests.
Every model call + tool dispatch + hook fire emits a phantom.* OTel span with latency, tokens, ok/error. Point OTEL_EXPORTER_OTLP_ENDPOINT at Honeycomb / Datadog / Jaeger. In-memory exporter for local debugging.
When you ask Phantom to build and run an app, it HTTP-GETs the URL after launch, detects 5xx / tracebacks in the body, reads the log, auto-patches the source, relaunches — before claiming "live". Tested against real Flask apps.
Multi-agent orchestrator runs dependency-ordered waves in parallel. SQLite-backed scratchpad lets upstream agents publish partial results downstream agents read mid-run — faster than re-running prompts. 6 built-in agent types with tool allowlists + worktree isolation.
/help /clear /model /memory /perm /hook /session /compact /cost /exit. Drop a ~/.phantom/commands/mycmd.md for your own templates with frontmatter description.
System → user → project-shared → project-local, with locked_keys for enterprise. 128K-token context compaction with PreCompact hook + LLM-backed summariser. Session save/load with atomic writes.
/web now does Claude-Code-style two-round research: round 1 scrapes the top 5 pages, the router model inspects the gaps and proposes 2 specific follow-up queries, round 2 scrapes 2 more pages each — ~9 pages total at 6,000 chars each. Output is clean narrative (no Step 1/2/3 spam, no inline citation noise). Run /sources after to see the URLs with first-line previews. ~30–60s wall time, dramatically deeper answers.
Follow-ups like "ui is not good" or "also add login" stay in the active project instead of spawning a new project_xxxxxxxx. Two-tier router detects refinement cues + active project and continues silently. Every refinement appended to phantom_summary.md as a run-history entry. Hardened multi-agent fallback: if the planner returns malformed JSON, a 4-agent default skeleton (Fetcher, Backend, Frontend, Runner) always produces a working build.
/update auto-exits the REPL on a successful upgrade so Python's sys.modules cache can't silently keep running stale code. Failed updates and "already-latest" cases stay in the REPL so you can retry. Backed by 4 dedicated tests + the full 775-Python-test suite (903 with TS + VS Code).
How PhantomCLI stacks up against other local AI agents.
Scroll table to compare →
| Feature | Phantom v1.0.2 | Claude Code | OpenClaw | AutoGPT | AgentZero |
|---|---|---|---|---|---|
| Any OpenAI-compatible model | |||||
| Local OS bash execution | |||||
| 9-event hook lifecycle | |||||
| Pattern permissions (bash:git:*, write:~/**) | |||||
| Hash-chained audit log | |||||
| CLAUDE.md hierarchy with live reload | |||||
| Multi-tier model routing (cheap/mid/expensive) | |||||
| VS Code extension with activity bar + code actions | |||||
| MCP resources + subscribe + reconnect | |||||
| Anthropic SSE + OpenAI SSE streaming | Anthropic only | ||||
| Programmatic Python hooks (@hook decorator) | shell only | ||||
| Parallel multi-agent with shared scratchpad | sequential only | ||||
| Self-heal after launching built apps | |||||
| OpenTelemetry spans + metrics | |||||
| Telegram two-way bot | |||||
| Persistent memory — episodic log + FTS5 RAG | |||||
| Image / Video / Voice APIs | |||||
| Permanent dangerous-cmd blocklist | |||||
| Live task-progress checklist | |||||
| Auto ML pipelines + dashboards | |||||
| One-time payment, lifetime | Anthropic sub |
"Getting local OS execution working was shockingly simple. It organised my entire Downloads folder in one sentence."
"Telegram integration is insane. I literally SSH into my VPS by texting my bot while commuting. Set it up in 10 minutes."
"Replaced 8 Python automation scripts with natural language. The God Mode terminator animation when you hit Level 4 is 🔥"
No monthly fees. No hidden charges. Buy once and own PhantomCLI for life.
localhost:8080 on your own machine. Aravind Labs servers are only involved in purchase and license key validation.irm "https://phantom.aravindlabs.tech/phantomcli/install.ps1?key=PHC-XXXX" | iexphantom serve in one terminal — it opens a unix-socket daemon. Run phantom connect in another (or any IDE/script) and you get sub-50 ms warm round-trip instead of paying Python startup every time. Tiered sandbox (Trust 1–4) and 48-pattern dangerous-command blocklist still apply. Trust 4 ("God Mode") still requires your licence key and auto-downgrades to Trust 3 after 30 min idle.phantom memory import claude-code (or codex, opencode) and your transcripts move into Phantom's episodic memory. phantom mcp import grabs ~/.claude/mcp.json and ~/.codex/mcp.json in one shot. No re-pasting server configs.python run.py setup → option 9. Need help? Email support@aravindlabs.tech.v1.0.2 ships today. Daemon, swarm, self-dev, voice, importer — all included. ₹999 once, yours forever, every future patch free.
Get Phantom v1.0.2 Now