Tech Plan — AI Shop Dashboard

SYSTEM VISION

Deploy 14 agent orchestration tools on a single machine — 7 boards, 7 CLI agents/harnesses — unified by a Slack/Mattermost communication layer, a shared memory system (SharedContext MCP + AGENTS.md), a local Ollama inference server, and a live dashboard with embedded terminal agents via ttyd. Every tool that discovers knowledge makes it available to every other tool in the stack.

14-Tool Inventory

7 orchestration boards + 7 CLI agents/harnesses on a single Apple M4 Pro machine

7 Orchestration Boards

#	Tool	Port	Purpose	Ollama	MCP
1	Claw-Kanban	4011	Lightweight task dispatch via # directives	Inherited	T1 native
2	Claw-Empire	4018	Strategic orchestration via $ CEO directives	Native	T1 native
3	Vibe Kanban	4012	Worktree-isolated parallel dev	No	Native
4	KaibanJS	4013	JS-native agent workflows	Indirect (proxy)	Native
5	AgentsBoard	4014	Jira-like AI task board	No	API-only
6	OpenClaw	4015	Multi-provider messaging bridge	Native	Bridge
7	NanoClaw	4016	Secure containerized agents (Claude-only)	No	T2 Claude.md

7 CLI Agents & Harnesses

#	Tool	ttyd Port	Role	Ollama
8	Droid	7687	Terminal coding agent (Factory.ai)	Native BYOK
9	Pi	7688	Minimal OSS coding agent toolkit	Native ext
10	Aider	7682	Git-native pair programmer	Native
11	Cline	7683	Autonomous plan-and-execute agent	Cloud API only
12	Crush	7684	Glamorous terminal agent (LSP, MCP)	Indirect (openai-compat)
13	Plandex	7685	Large-task planner with diff sandbox	Self-host
14	Amp	7686	Frontier coding agent (cloud-only)	Not available

Local Ollama Model Stack

Apple M4 Pro · 64 GB unified memory · Budget: model file ≤38–45 GB

Model	Size	Speed	Role
qwen3-coder:30b	19 GB	~64 tok/s	Default coding All tools
qwen3:32b	~20 GB	~15–22 tok/s	General reasoning, planning
deepseek-r1:14b	~9 GB	Fast	Quick feedback, sanity checks
qwen2.5-coder:32b	~20 GB	~15 tok/s	Cline local profile
devstral:24b	~15 GB	~20 tok/s	Plandex secondary

Model Profiles (config/model-profiles.yml)

balanced

primary: qwen3-coder:30b · fallback: deepseek-r1:14b

fast

primary: qwen3:8b · fallback: deepseek-r1:8b

quality_local

primary: qwen2.5-coder:32b · fallback: qwen3-coder:30b

plandex_local

primary: qwen3:32b · fallback: devstral:24b

DO NOT RUN on M4 Pro: qwen3-coder:480b, deepseek-r1:70b, gpt-oss:120b — these exceed the 38–45 GB model budget.

Memory Architecture

Two complementary layers: AGENTS.md (declarative) + SharedContext MCP (dynamic, encrypted)

AGENTS.md (Filesystem)

Declarative, low-churn · Read by 12/14 tools

Project conventions + code style
Routing directives (# and $ patterns)
Active constraints + merge freezes
Cross-agent handoff notes
Tool-specific sections per agent

Location

project root / AGENTS.md

SharedContext MCP (SQLite)

Dynamic, structured · Read by 9/14 tools · AES-256-GCM encrypted

Semantic fact search across sessions
Conversation handoff between tools
Namespace-scoped isolation
High-frequency dynamic facts

MCP Tools Exposed

store_fact recall_facts list_facts delete_fact store_conversation recall_conversation

Memory Namespaces

project:<name>

Bug patterns, arch decisions

tool:<name>

Tool-specific preferences

global

Cross-project conventions

session:<id>

Conversation history for handoff

slack:<channel>

Thread context

MCP Connectivity by Tool

Connection Type	Tools
Native MCP client	Claude Code, Cline, Crush, Amp, Droid, Vibe Kanban
Inherited from host CLI	Claw-Kanban, Claw-Empire, NanoClaw
Bridge adapter needed	Aider (CLI wrapper), Plandex (context inject), Pi (extension), OpenClaw (skill), KaibanJS (JS adapter)
Not practical	AgentsBoard (API-only, no extension system)

Communication Layer

Slack (human ↔ bot) + Mattermost (bot ↔ bot) dual-channel architecture

Slack Bridge (Primary)

@slack/bolt · Socket Mode (no ngrok) · Port 4010 · SQLite thread persistence

Routing Rules:

Channel message → look up slack-project-map.yml → route to project
DM / Group DM → parse project:<name> → route
Board mention (@claw-kanban) → route to board API
Tool mention (tool:aider) → spawn CLI agent
Thread continuity → channel_id + thread_ts → session in SQLite
Fallback → help message listing all projects and boards

Bot scopes: chat:write, channels:history, groups:history, im:history, mpim:history · Safety: ignore bots, rate-limit 10 req/min/channel

Mattermost (Secondary · Bot ↔ Bot)

Team Edition in Docker + Postgres · Port 4020 · Bridge port 4021

Channel Layout:

#agent-bus

Inter-agent broadcast (all 14 tools)

#tool-claude, #tool-aider, etc.

Per-tool activity channels

#project-<name>

Per-project cross-tool coordination

POST /agent-bus

{ from_tool, message, project, priority } → broadcasts to #agent-bus

6-Phase Implementation Sequence

Run make <target> at each step — all phases scripted in the Makefile

1

Prerequisites

Node.js ≥22.16.0 (OpenClaw enforces)
pnpm, npx, python3, pip, git, docker, ttyd
All CLI agents: claude, aider, cline, crush, plandex, amp, droid
Ollama: ollama --version
Smoke test each CLI: <tool> --version

2

Install Tools

Clone all 7 boards (git clone → pnpm install)
Pin Claw-Empire v2.0.4, OpenClaw 2026.3.13
Install CLI harnesses (pip, npm, brew, curl)
Ollama: pull qwen3-coder:30b (19 GB), qwen3:32b, deepseek-r1:14b
Slack bridge: Bolt for JS + SQLite

3

Shared Memory

Clone SharedContext MCP → npm install → build
Generate recovery phrase → .env SHARED_MEMORY_RECOVERY_PHRASE
Configure MCP-native tools (Claude, Crush, Droid, Amp, Cline, Vibe)
Set up bridge adapters for Aider, Plandex, Pi
Create AGENTS.md template + sync script

4

Mattermost

docker-compose.yml (Team Edition + Postgres)
make mattermost-up → start containers
setup-mattermost.sh → admin + agent-hub bot + channels
Create mattermost-bridge/app.js (webhook router)
Configure outgoing webhook → bridge

5

Dashboard & Terminals

brew install ttyd
Express dashboard: launcher/server.js on :4000
Health grid: card per service, auto-refresh 10s, green/red badge
make ttyd-up → 8 ttyd instances (ports 7681–7688)
Embedded terminal iframes for all 8 CLI agents

6

Tests

Smoke: CLI version checks + provider adapters
Service boot: verify all ports 4010–4018, 7681–7688
Slack routing: channel → project, DM, thread continuity
Integration: cross-tool memory (Claude stores → Crush recalls)
Failure paths: missing keys, port conflicts, Docker unavailable

Key Makefile Commands

Setup Commands

make check-prereqsNode, CLIs, docker, ollama, ttyd

make installClone/scaffold all 14 tools

make ollama-setupPull recommended local models

make memory-setupInstall SharedContext MCP

make slack-setupApply Slack app manifest

make mattermost-setupStart MM + create bot/channels

Run / Test Commands

make startLaunch everything

make statusRunning services + ports

make ttyd-upLaunch 8 ttyd instances

make dashboard-upExpress dashboard on :4000

make testAll smoke + routing + integration

make memory-statusFact count per namespace, DB size

Key Design Decisions & Risks

Decision	Rationale
Bolt bridge is primary Slack entry (not OpenClaw)	OpenClaw daily releases can break routing. Bolt bridge is stable, single-purpose. OpenClaw's Slack adapter is optional complement.
Both Claw-Kanban AND Claw-Empire installed	Complementary roles: Kanban = fast CLI dispatch, Empire = full orchestration with API/OAuth/messenger.
SharedContext MCP (not OpenMemory)	SQLite-based, no Docker dependency, encrypted, lightweight. OpenMemory requires Postgres + Qdrant.
ttyd (not xterm.js + custom backend)	Zero-code, one command per agent, embeddable iframe. xterm.js requires writing WebSocket/PTY server.
Node ≥22.16.0 (not ≥18)	OpenClaw enforces at startup. Required.
AGENTS.md + SharedContext dual-layer	AGENTS.md is universally readable (12/14 tools). SharedContext handles high-frequency dynamic facts. Neither alone is sufficient.

24 Acceptance Criteria

Core (1–4)

make check-prereqs passes (Node ≥22.16.0)
make install completes without manual intervention
make start runs all 14 tools + bridges + Ollama + dashboard
Dashboard shows health grid with working links at :4000

Boards + Agents (5–10)

Claw-Kanban responds on port 4011
Claw-Empire responds on 4018, CLI agent detected
All 7 boards pass boot tests on assigned ports
Each CLI agent responds to a trivial prompt
Droid, Pi, Aider, Crush, Plandex respond via Ollama
Amp responds via Sourcegraph free tier

Comms + Memory (11–24)

Slack channel messages route to correct project
Slack DMs route via project:<name> prefix
Thread context preserved across follow-ups
Mattermost starts on :4020, bot created
Agent-bus broadcasts to all subscribed tools
SharedContext MCP starts, memory.db created
Cross-tool memory: Claude stores → Crush recalls
Namespace isolation verified
Claw-Kanban spawned agent inherits SharedContext
AGENTS.md exists with all required sections
/api/health returns all services status
ttyd iframes are interactive in dashboard
Health grid auto-refreshes every 10s
make test passes all smoke, routing, integration, memory

Final Tech Plan