neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
翻譯準備中
SpecWithCode: free Mac app for Claude Code, Codex, Copilot and other coding agents. Download the macOS .dmg at spec-weave.com/code/download
翻譯準備中
取自儲存庫 README
正在載入完整 README。
檢視完整 README ↗Write the spec. Get the code.
SpecWithCode is a free desktop workspace for the coding agents you already pay for. Run Claude Code, Codex, Grok Build, Cursor, OpenCode, Copilot, Antigravity and Pi side by side on your own machines. Every change an agent makes lands with its spec: what was asked, what counts as done and the proof, in the same repo as the code, so any agent on any machine can pick it up.
It is built on SpecWeave, the open-source CLI (npm install -g specweave) that keeps that record in git. Website: spec-weave.com/studio.
Built for teams of 2–10 people. Free, no account. You bring your own provider subscriptions or API keys; SpecWithCode supplies no inference, resells no tokens and charges no seat fee to run agents on your team's machines. What is free today stays free; teams can pay for hosted private skills and support (pricing).
Automatic fallback requires an explicit configured choice before selecting ACP agents whose billing is unknown. See when a provider runs out.
HTML reports open from a compact preview into a focused reader. The conversation stays available through Show thread. See the report experience for behavior and compatibility limits.
Settings > Skills manages versions and agent visibility. Claude plugin skills use the whole-plugin control; disabled plugins stay visible so you can turn them back on or Undo. Version updates require a fresh preview if the registry changes before confirmation.
| Instead of | In SpecWithCode |
|---|---|
| Claude Code Projects, Codex cloud tasks | Projects: a coordinator splits a goal into tasks and hands them to workers, each on its own provider and model. Memory, a plan and verified tasks per project. |
| One app per lab | Every official agent in one window, each with its own sign-in, approvals and limits. Switch a thread to another provider mid-conversation and the recent conversation carries over. |
| Waiting for a usage reset | Usage limits for every account on one screen. When a provider runs out, a per-project rule offers another provider, waits for the reset or switches on its own. |
| OpenAI Dots, Hermes, OpenClaw | Personal projects with their own folder, memory and scheduled routines; local MCP tools granted one by one; a Telegram channel to the coordinator through your own bot; optional read-only Gmail, Calendar and Drive. |
| A cloud container | Your machines. Work runs where your code, skills and folders already are. Pair more machines over your local network, an SSH tunnel or Tailscale. |
| Memory locked in one vendor | The record in git. Specs, task evidence and handoffs are files any tool can continue from with specweave pickup. |
SpecWithCode is not a cloud service: work runs on the machine you choose and stops when it sleeps.
Settings → Connections includes a read-only MCP connections library for Codex and Claude Code on each machine. Project Environment adds project declarations and overrides alongside SpecWithCode-managed tool grants. Configuration is shown separately from unverified authentication and session availability.
The model picker distinguishes the coding harness from the model and its billing source. See the harness integration guide for OpenCode, Pi, native adapters and verification limits.
SHA256SUMS-windows-linux. Intel Macs are not built.The app updates itself from this repository's releases. Turn that off in Settings → About, or set SPECWITHCODE_DISABLE_AUTO_UPDATE=true (the older STUDIO_DISABLE_AUTO_UPDATE is still accepted).
More: install, first run, providers, remote access, outside agents over MCP, your own VPS, updating.
In Settings > Providers, a CLI's update action opens a machine selection plan. Choose only this CLI here, or explicitly include eligible connected machines. The app uses each configured executable's owning installer and reports each outcome; see provider updates.
To connect Copilot, Grok or Gemini, use Add or manage providers in the model picker. Provider setup explains installation, sign-in and model discovery; a connected usage source only reports quota.
The composer’s access selector also controls a provider’s Allow All setting. Full access enables it; all other access modes disable it. Agent mode and reasoning remain in the model-options menu. If the provider cannot confirm the permission change, the turn stops with an error.
Start a thread: pick a folder, a provider and a model, and talk to the agent. Plans, approvals, steering, queued turns and git checkpoints work as in the agent's own app.
Start a project when the work is bigger than one thread: give it a goal and let the coordinator hand tasks to workers. See project orchestration.
Add a machine: run SpecWithCode on it, then pair it. Being reachable is not permission; every machine and browser is paired separately. To pair a browser:
specwithcode auth pairing create --ttl 5m --label "My laptop" --base-url http://127.0.0.1:8319
The output is a credential. Keep it out of chat and screenshots.
The local service listens on http://127.0.0.1:8319/ and keeps its data in ~/.specwithcode; installs that started before that default keep using ~/.specweave-studio. The command line tool is specwithcode, also installed under its original name specweave-studio.
Projects keep their coordinator in the main pane, with workers beside it. For short replies, use Settings → Agents → Response style → Concise. If the coordinator's model is unavailable, open its model picker to Check again, choose another model, or open setup for that machine; checking availability does not resend your message.
The unreleased project handoff model keeps one named project and coordinator window across explicitly linked machines. Threads retain their execution machine and recorded model; earlier coordinator segments remain readable. Handoffs start a fresh provider session, and unreachable machines can have history that is temporarily unavailable. Existing independent projects require explicit association.
Folders, sessions and project data stay on your machine. Telemetry and the hosted relay are off by default. Model calls go to the providers you choose, as they would from your terminal. See security boundaries and telemetry.
Node 22.20 or later and pnpm 11.
pnpm install --frozen-lockfile
pnpm dev # server and web app
pnpm dev:desktop # desktop app
For app-only work (most lanes), pnpm run install:lane installs everything except
infra/relay — it skips the Alchemy/Cloudflare deploy toolchain (≈2 GiB of packages).
Add the relay later with pnpm run install:relay. In an existing full checkout, run
pnpm run clean first: pnpm will not shrink an already-linked tree on a scope switch.
Release and CI always run the full install.
Contributor notes, architecture and focused checks are in AGENTS.md.
Early access. Projects, providers, usage limits, machines, routines and the Telegram and Google connections are in the current build and in daily use. Live project runs are recorded with Codex and Claude Code; Grok Build and OpenCode have tutorial runs; Cursor and Antigravity workers have no recorded project run yet. See the changelog.
Source is available for inspection. No new project-wide license is granted for original SpecWithCode material; previously distributed versions retain their existing grants. Third-party code and assets retain their own licenses and required notices in THIRD_PARTY_NOTICES.md.