Phantom Phantom · comparison
When one workbench beats five terminals.
The coding-agent CLIs are good at being CLIs. Phantom does not replace them: each turn still spawns the vendor's own binary under the account you chose. What it adds is the layer around them. Vendor statements below link to the vendor's own public pages; Phantom statements link to this repository's docs and source.
What the tools already are
Credit where it is due.
Each vendor ships its own agent interface with its own update cycle. These statements paraphrase the linked public page; where no public source exists, we say nothing.
Claude Code
Anthropic's agentic coding tool: reads your codebase, edits files, runs commands, and is available in the terminal, IDE extensions, a desktop app and the browser.
Codex
OpenAI's lightweight coding agent that runs in your terminal, published as open source under the Apache-2.0 license.
Source: github.com/openai/codex
Grok CLI
xAI documents headless CLI scripting for Grok Build (grok -p style non-interactive runs) alongside sandbox limits in its own docs.
Source: xAI docs, Grok Build scripting
Devin CLI
Cognition's CLI for Devin, powered by the SWE-2 model family; cloud sessions are a separate API-driven product with their own pricing.
Source: devin.ai/cli, devin.ai/pricing
Ollama
Serves an OpenAI-compatible API at localhost:11434/v1 for the models you pull, per its own compatibility documentation.
Source: Ollama OpenAI compatibility
What a workbench adds
The layer around the terminals.
Each claim cites the repository file or guide it comes from. Phantom ships no model of its own and keeps the vendor harness intact: the turn still runs the real CLI.
One console for every seat
Chats, fleet work, resources and evidence in one window. Each turn spawns the vendor CLI process in the project directory and resumes its conversation by id, so the memory stays the vendor's own.
Source: docs/VERSE.md
Pinned profiles per account
Multiple accounts at the same provider stay in separate prepared profiles, so seat choice is an explicit decision rather than whichever login the CLI last used.
Source: docs/VERSE.md
A context meter that reads the real window
Windows and compaction points come from the pinned binaries and each seat's catalog, not provider marketing pages. Warnings fire at 80% and 95% of the compaction point.
Source: docs/VERSE-CONTEXT.md
Reset-aware seat choice
When several seats could do the work, the fleet can prefer an eligible account whose qualified allowance finishes before its deadline. Missing evidence stays unknown.
Source: docs/RESET-AWARE-SCHEDULING.md
Review as a first-class surface
Per-turn checkpoints, per-file and per-hunk accept and reject, undo and rewind with previews, and a recorded decision log. The raw CLI gives you the diff; the workbench gives you the controls.
Optional signed autonomy
On macOS, a Touch-ID-signed standing grant can run fleet work unattended inside its scope: an eight-stage ladder that starts in shadow, fixed merge gates and reversible merges. Dormant until you set it up.
Source: docs/STANDING-AUTHORITY.md, docs/RESIDENT-RUNTIME.md
The honest other side
Drive the CLI directly when it is enough.
One seat, one-off work
If you use a single account for a single task, the vendor CLI alone is the shortest path. Phantom's value shows up when seats, projects or fleet work multiply.
Non-interactive pipelines
The same headless invocations Phantom drives (claude -p, codex exec, grok -p) work fine in a CI script without a console. Phantom does not need to sit in that path.
Minimal footprint
Phantom is a Node.js 22.15+ install plus the vendor CLIs it drives. If you do not want the console, the checkpoint store or the optional daemon, keep the CLIs and skip it.
Source: docs/QUICKSTART.md
Keep going