neuron-explorer
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
An unofficial MCP server and CLI that let AI agents and scripts read and wait on Better Agent Terminal sessions over BAT's remote protocol, and nudge or orchestrate them only on hosts where you enable
개발 도구Python · JavaScriptGitHub @TED-H-com온라인운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
Unofficial MCP server and CLI that let AI agents read, nudge, and orchestrate Better Agent Terminal sessions
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
An unofficial connector that lets AI agents work with Better Agent Terminal (BAT) sessions.
BAT (by TonyQ / tony1223) is a terminal app that runs Claude Code and Codex agent
sessions, grouped into workspaces, on your machines. It has a remote protocol (bat-remote/v2) that its own GUI and
phone clients use. This project speaks that protocol so that other agents (Claude Code, Codex, Cursor, Hermes, or any
MCP client) and shell scripts can:README 전문 보기 ↗Understand the project. Discuss the problem. Decide with the PM.
English · 繁體中文 · Website · Install and start · Progress
Better Agent Dashboard keeps project state, PM discussions and decisions in one workspace. Start with the live Overview, open a problem's original record, and discuss it with the PM without rebuilding the context. Review the proposed update and explicitly adopt it; requirements, work and delivery keep their own evidence.
Open the interactive Dashboard demo.
Explore the actual shared interface with synthetic, read-only project data: search work,
expand requirements, inspect delivery evidence and browse sessions. The demo also links to
the v1.3 workflow preview, which simulates the proposed start, follow-up and decision flow.
Neither example connects to a BAT host or model.
The real Start work panel is on main and needs a configured BAT host and a bound repository.
Published preview source 437ee6a7 (through #395, October 22, 2026) connects reviewed plans to Work, queued follow-ups, fixed Results, verification and PR delivery. Workbench provides mobile navigation and foreground recovery; profiles show route limits before start. PM can discuss two fixed assessments, save continuation briefs and explicitly assess the Overview. Resources reads separately authorized local Codex/Claude histories. Lists load summaries first; unsent instructions can return to Idea. Core UI actions share HTTP/MCP contracts. Classic remains the default. Integration fixtures use real Connector and disk Git with mock BAT/providers; live BAT/model and physical-device acceptance remain separate.
Actual shared UI · synthetic demonstration data · merged-main preview · capture source.
The Web and Tauri desktop apps share one interface and one Python backend. You keep using Better Agent Terminal (BAT) for your own coding. Windows Fleet manages local connections and startup. Connector keeps project history, permissions and operations consistent across the Dashboard, CLI and MCP.
Development preview. The screenshots come from the published prerelease source
437ee6a7(October 22, 2026, through #395). The gallery uses synthetic data; release, installation and live acceptance have separate evidence. The progress report separates branch, merged, installed and live-tested status. Match a successful package to its exact source; screenshots and CI do not establish native or live acceptance.
| Your question | View and workflow |
|---|---|
| Where does the project stand? | Overview shows the live goal, accepted requirements, blockers, open decisions and Recently decided history. Needs attention lists each record with one Discuss action. Unread sources stay Unknown; complete empty lists say None recorded. Saved reports retain their fixed snapshots. |
| What needs a decision? | Open a record and Discuss with PM. Discussions opens with its context chip and automatic title. Search threads, save a note, and review a proposed update. Context, Inputs and Plan open in Discussion info. An adopted discussion can offer an authorized, editable Resolve issue step; reported resolution remains unassessed. |
| What is this work meant to satisfy? | Requirements shows adopted requirements with work, PR, Acceptance and Deploy columns. Expand a row for criteria and work; open the inspector for evidence. Unchanged refreshes keep rows and focus stable. |
| What is still unresolved? | Issues & changes keeps risks, issues, actions, changes and decisions in a filterable register with original timelines and named maintenance findings. |
| Which dates changed? | Schedule compares recorded baseline, forecast and actual dates on a labelled axis. Missing dates stay unknown. |
| What could a dependency change affect? | Dependencies shows pinned inventories, evidence and bounded impact in a matrix or graph. |
| What are the agents using? | Agents keeps sourced usage, rows and daily trends on one consistent read, with reported quota windows. Unavailable data stays unknown. |
| What should happen next? | Board offers search, requirement and agent filters, plus the shared work inspector. When the workspace can start work, that inspector includes Start work and, for a started attempt, a read-only Result section. Content approval, requirement acceptance, merge and deployment remain separate. |
| What needs a recorded response? | Open Governance or Follow-ups from Issues & changes. Meeting threads have a Meeting tab in Discussion info. Saved outcomes keep their original sources. |
| What may maintenance do? | Dependencies holds Release baselines and . Preview a release source before registration; policy adoption and enabled repair limits remain explicit. |
These are views of the same project, with secondary tools available on demand. The current scope distinguishes the merged workspace from the full PM v1.2 backlog. The progress report lists work still in review.
A discussion does not automatically adopt a plan; adoption does not start work or resolve its source issue. After adoption, an authorized person can explicitly report the issue resolved; this does not accept a requirement. Policy-bounded delivery and maintenance remain available through the project's secondary tools. Step-by-step PM guide.
| More views from the same synthetic project | Screenshot |
|---|---|
| Issues & changes | Register with filters |
| Schedule | Baseline, forecast and actual |
| Dependencies | Pinned partial inventory graph |
| Board | Work lanes with search and filters |
| Session | Read-only manual session; host not set up for managed work |
The latest stable download is Connector 0.2.12 / desktop 0.1.8, pinned to 14c62830. The published prerelease is Connector 0.2.13 / desktop 0.1.9, pinned to 437ee6a7143a82e32eae1d6a008203b2df072aaf (437ee6a7). Each release has eight assets: a Python wheel and sdist, a Linux deb, two macOS DMGs, a Windows installer, release-manifest.json and SHA256SUMS. Choose the desktop asset on the release page and check its manifest and checksums. These remain unsigned or ad-hoc-signed development builds, not a stable signed update channel.
For Connector, follow the verified installer instructions. Use --tag v0.2.12 --channel-check stable for stable, or --tag v0.2.13-preview.20261022 --channel-check preview for this preview, with the bootstrap from the same tag. The installer requires a tag; its default channel check is preview, which accepts stable releases and prereleases. stable refuses a preview. On Linux, --enable-updates enrolls and schedules that same channel; without it, installation alone does not enable updates. See the installer source and update guide.
The enrolled updater offers eligible newer stable releases on stable, and eligible newer stable releases or prereleases on preview. It does not offer this preview to the stable channel. For later development builds, match a successful desktop workflow on main to its source; Actions artifacts can expire.
| Entry | Package / local features | Boundary |
|---|---|---|
| Windows x64 | desktop-Windows-unsigned: NSIS, bundled central, tray, Credential Manager and Fleet | Unsigned validation package. Windows Fleet manages connections and BAT startup. |
| macOS Apple Silicon / Intel | desktop-macOS-arm64-validation / desktop-macOS-x64-validation: DMG, bundled central, menu bar and Keychain | Ad-hoc signing; Windows Fleet is not ported to Mac. |
| Linux | desktop-Linux-unsigned: Debian package and bundled central | External-central native credentials remain memory-only; persistent native vault is excluded. |
| Browser | Open the central's /dashboard/, or use the desktop's browser entry | Same project data and UI; no browser-native Fleet or credential vault. |
The desktop package prepares a matching Python runtime, private background central, data and personal identity.
Connector automatic updates are available for verified per-user Linux installations. Once enrolled, the hourly timer installs eligible releases for future launches; existing MCP, central and BAT processes keep running.
First use opens Get ready to work → 1. Connect BAT. Paste BAT’s connection URL, check the Host ID, then Verify and save host. After success, choose View sessions. Repository binding and verification commands are optional collapsed steps; configure them when the work needs them.
For managed work, keep BATC_TASK_SETTINGS private with mode 0600. Configure [task_service.base_branches] per project, [verification.ssh_hosts] per task host (required when using configured base branches), and [verification.commands] for trusted tests.
Joining an existing central is an explicit advanced option. A failed connection never creates a replacement project database. Closing the Dashboard window does not stop central work.
PM analysis has separate runtime requirements. The current bounded analysis paths run on a configured Linux central or eligible Linux BAT host, using the pinned PM Codex runtime and confinement described in the guide. A Windows or Mac client can use those remote capabilities; its local desktop installation does not provide native Windows/Mac PM execution. PM analysis support is distinct from a general coding CLI adapter.
Installation and first use · Windows guide and resources · Recovery and upgrades
flowchart TB
Web[Web Dashboard] --> Central[Python Connector / Task Service]
Desktop[Tauri Dashboard] --> Central
API[CLI / MCP / agent integrations] --> Central
Desktop --> Fleet[Windows Fleet: connections and startup]
Fleet --> Routes[SSH routes / selected BAT profiles]
Routes --> BAT[BAT hosts and workspaces]
Human[Your BAT desktop / mobile] --> BAT
Central --> Journal[Plans / evidence / permissions / operations]
Central --> BAT
BAT --> Manual[Human resources: Connector read-only]
BAT --> Managed[Agent-owned sessions and worktrees]
Managed --> Delivery[Fixed commits / PR / merge / deployment]
Delivery --> Central
desktop/src; drafts and reading positions remain local to each client. Hermes, Grokbot and other MCP/CLI clients use their own principals and scopes.| Status | Scope |
|---|---|
| Merged baseline | Desktop/central foundation; PM discussions and plan adoption; engineering packages and bounded dispatch; verification, requirement acceptance and dependency evidence; policy-bounded maintenance repairs; BAT proposals/Follow-up; AHP read-only observation. #132 adds central Overview, discussion context and renames, register, Schedule, Dependencies and daily resource buckets. #134 adds adopted goals, recent decisions and the guided close of a resolved issue. |
| Merged workspace | Published preview source 437ee6a7 (through #395, October 22, 2026) connects reviewed plans to Work, queued follow-ups, fixed Results, verification and PR delivery. Workbench provides mobile navigation and foreground recovery; profiles show route limits before start. PM can discuss two fixed assessments, save continuation briefs and explicitly assess the Overview. Resources reads separately authorized local Codex/Claude histories. Lists load summaries first; unsent instructions can return to Idea. Core UI actions share HTTP/MCP contracts. Classic remains the default. Integration fixtures use real Connector and disk Git with mock BAT/providers; live BAT/model and physical-device acceptance remain separate. |
| Remaining PM v1.2 | General native coding adapter, AHP create-own/write support, remaining document views and presentations, wider CLI support, remote usage collection and delivery-efficiency attribution. |
Local Codex/Claude usage collection exists; the redesign's Agents view adds sourced daily trends and reported quota windows. Remote collection and delivery-efficiency attribution remain planned. Missing usage stays unknown; observed tokens, API price estimates and subscription quota are different measures. No usage-driven model policy is active. Progress and priorities.
Formal signing, Mac notarization, production update channels and Linux persistent native credential storage are outside the agreed delivery. Real-environment acceptance belongs to the owner and has not been claimed as passed. The full PM v1.2 plan remains the scope; M4 is one milestone.
Human sessions and worktrees remain read-only through Connector. Unknown resources need original creation evidence before managed recovery. Reading another session never transfers control of it.
A lost reply does not prove an action failed. Recover the original operation and key; uncertain work keeps its evidence and reservations. Current permissions and source availability are checked before effects. Connector policy complements host accounts and confinement; it is not an operating-system sandbox.
Security model · Resource policy · Operations
| Need | Start here |
|---|---|
| Use the product | First use · Windows Fleet |
| Understand progress | Progress report · PM status · October 17 report |
| Continue development | Documentation handoff · Adopted plan · Architecture decisions |
| Integrate an agent | Agent skills · Automation API · API · Protocol |
| Follow results | Artifacts · · |
uv sync --locked --extra dev
uv run ruff check .
pytest-disktmp uv run pytest -q
cd desktop
npm ci
npm run build:all
npm test
npm run test:ui
Use a disk-backed pytest temporary directory and remove it after the run. pytest-disktmp is the development host wrapper; if unavailable, supply a unique disk-backed --basetemp. Never use /dev/shm or another tmpfs. Backend changes also require the Python 3.10 suite. Follow CONTRIBUTING and AGENTS.md.
Edit the shared UI in desktop/src and regenerate browser assets. The public website is generated from repository source. Tests use mock BAT/providers and temporary repositories, without live-host writes.
An unofficial MIT-licensed companion to BAT, not affiliated with or endorsed by its authors. BAT is by TonyQ and contributors. Selected organization and interaction patterns reference Project Hub; no Hub importer or runtime dependency is included. See the adoption review.
| Where are the work tools? | Requirements → More actions opens quality, evidence, specification, package, handoff and issue publication tools; Board → More actions opens Work batches. |
| What happened in that session? | Session keeps the conversation and results together. Authorized managed controls are available; manual BAT sessions remain read-only. |