Repository navigation
Releases
Weft v0.1.12 — live activity + instant Gemini message
Universal macOS build (Intel + Apple Silicon).
New
- Live activity — the model's thinking and every tool step it takes (📄 Read, ✏️ Edited, ✍️ Wrote, ⚙️ Ran, 🔎 Searched, 🤖 Sub-task…) now stream live as it works, growing in place, then collapse to “Thought for Xs”. No more silent wait until the end. The answer stays clean.
- Gemini responds instantly — since Antigravity/agy has no working headless mode (Google deprecated the standalone gemini CLI; agy
--printtimes out), Weft now shows the clear “use it in the terminal (esc →agy); Claude & Codex work here” message immediately, instead of hanging ~30s.
Also in recent releases
- Warm sessions (fast after the first message), real file editing, full/warm mode default (claude-mem memory + skills + MCP).
- Token line (
🔢 in · out) in the footer under the reply. - Codex: clean, actionable message when your ChatGPT plan rejects the model.
- No code-sign crash on update; stable signing (keychain stops re-prompting).
Install (no password, no Gatekeeper block)
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-mac.sh | bash
Claude + Codex work in the agent. Gemini works in the terminal.
Weft v0.1.11 — warm sessions, visible activity, graceful CLIs
Universal macOS build (Intel + Apple Silicon).
New
- Warm sessions — one live
claudeprocess per chat: the first message loads claude-mem memory + skills once, every message after is fast. - See everything it does — a collapsible “Thought & worked” dropdown shows the files it reads/edits/writes and commands it runs; the answer stays clean.
- Real editing — the agent actually edits files / runs tasks in the folder you opened.
- Full mode is default — claude-mem memory + skills + MCP; type
fastfor stripped, memoryless speed.
Fixes
- Token line (
🔢 in · out · cached) now sits in the footer under the reply, not above it. - agy/Gemini: stopped sending a flag Antigravity rejects; and since Google deprecated the standalone gemini CLI and agy has no working headless mode, Weft now shows a clear “use it in the terminal” message instead of hanging.
- Codex: clean, actionable message when your ChatGPT plan rejects the model (instead of a raw 400).
- No more code-sign crash on update — the installer/updater quits a running Weft before replacing it.
- Stable signing identity — macOS keychain stops re-prompting on every update.
Install (no password, no Gatekeeper block)
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-mac.sh | bash
Claude + Codex work in the agent. Gemini (Antigravity/agy) works in the terminal.
Weft v0.1.10 — clean streaming, numbered models
Fixes the big one: streaming duplication. Replies now render as ONE clean block instead of stacking repeated copies (the 'pyramid'). Verified live.
Fixed
- No more duplicated/stacked responses — the live stream now emits each block once.
- Every model shows a version number — the Claude list is versioned-only (
claude-opus-4-8,claude-fable-5, …); the vague 'default' row is gone.
Universal macOS (Intel + Apple Silicon).
Install (no Gatekeeper warning)
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-mac.sh | bash
No login, no account, no API keys — runs your own Claude / Codex / Gemini CLI subscriptions.
Weft v0.1.8 — model prices, Universal Mac
Universal macOS build (Intel + Apple Silicon, one .dmg).
New in this release
- $ per 1M tokens shown for each model in the /MODEL picker (e.g. Opus $15/$75, Gemini Flash $0.30/$2.50), under the Intelligence / Speed / Cost bars.
- Everything from v0.1.7: MCP servers section + real model spec bars.
Install
- macOS (no security warning):
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-mac.sh | bash— or download Weft.dmg below. - Linux:
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-linux.sh | bash - Windows (PowerShell):
irm https://raw.githubusercontent.com/black12-ag/weft/master/install-windows.ps1 | iex
No login, no account, no API keys — runs your own Claude / Codex / Gemini CLI subscriptions.
Weft v0.1.7 — Universal Mac, MCP section, model specs
Universal macOS build — one .dmg runs on both Intel and Apple Silicon Macs.
New
- MCP servers in Settings → Third party CLI agents — see each CLI's MCP servers with on/off toggles.
- Model specs — the /MODEL picker now shows real Intelligence / Speed / Cost bars per model (no more "?").
- No-warning install —
curl -fsSL https://raw.githubusercontent.com/black12-ag/weft/master/install-mac.sh | bashinstalls with no Gatekeeper warning.
Install
- macOS: download Weft.dmg below, or run the install command above.
- Build from source: Linux
install-linux.sh, Windowsinstall-windows.ps1.
No login, no account, no API keys — runs your own Claude/Codex/Gemini CLI subscriptions.
Weft v0.1.6 — clearer UX, more CLIs, in-app links
New in v0.1.6
- Clearer chat input — empty box now says
Ask anything, e.g. …and the tip reads "Type your message and press Enter to chat." No more cryptic ⌘I toggle to figure out. - 7 CLIs in Settings — Settings → Third party CLI agents now lists Claude, Codex, Gemini, OpenCode, Cursor, Copilot, Ollama, each with a Connect / Disconnect button that runs the CLI's own login.
- Picker shows only what's active — the model list shows models from CLIs you're signed into; the full menu (with Connect) lives in Settings.
- In-app links — the welcome screen has a footer: ⭐ Star on GitHub · Portfolio · WhatsApp · Telegram.
- "New Weft conversation" header (rebranded).
- Carries the earlier Gemini fix, Codex effort ranges, and
~/.weft/clis.tomlfor adding any other CLI.
Install
Download Weft.dmg, drag Weft to Applications, right-click → Open (once). Connect a CLI in Settings, pick a model, and chat.
Weft v0.1.2 — more CLIs, Gemini fix
New in v0.1.2
- Gemini fixed — the
--print took --model as its prompterror is gone. - More CLIs, auto-detected — Weft now finds and lists every installed CLI in the model picker, not just Claude/Codex/Gemini:
- OpenCode, Cursor (
cursor-agent), GitHub Copilot - Ollama (local) — your installed models, pulled from
ollama list, run offline
- OpenCode, Cursor (
- Add any other CLI via
~/.weft/clis.toml— declare name + binary + prompt flag and it appears (no code change). Example in the README. - Codex effort ranges corrected per model family (no more
minimalrejection).
Extras that need sign-in: run their own login once (opencode auth login, cursor-agent login); Ollama needs nothing.
Install
Download Weft.dmg, drag Weft to Applications, right-click → Open (once). Then connect a CLI and pick a model.
Fork of Warp OSS. MIT/AGPL.
Weft v0.1.0 — CLI subscriptions in one terminal
Weft runs your CLI subscriptions (Claude Code, OpenAI Codex, Google Gemini) in one terminal chat — no account, no login, no API keys needed.
Install
- Download
Weft.dmgbelow. - Open it, drag Weft to Applications.
- First launch: right-click Weft → Open (it's ad-hoc signed, so macOS asks once). Or run:
xattr -dr com.apple.quarantine /Applications/Weft.app - Install/sign in to at least one CLI:
claude auth login·codex login· rungeminionce. (API key instead of a subscription?codex login --with-api-key, or setANTHROPIC_API_KEY.) - Open Weft → pick a model from the model button → go.
What's in it
- All models from each CLI, with per-model thinking levels
- Live-streamed answers + collapsible thinking + token counts
- Switch models mid-chat, shared cross-provider memory (Obsidian brain)
- Fast mode (Claude ~4s, Codex ~6s);
tools onfor full MCP/skills - Agentic Codex (reads files, runs commands in your project)
- Settings → Third party CLI agents: Connect/Disconnect each CLI
A fork of Warp OSS — Warp built the terminal/UI/UX/backend; Weft changes the AI layer to run locally. Licensed MIT/AGPL.