Agent Middleware

AI-native task orchestration. Any agent plans, middleware executes.

Agent → Plan (DAG) → Middleware → Workers (AI / API / Shell / Logic) → Results

Features

• DAG Plan Engine — Streaming dispatch, step-level retry, convergence loops
• 7 Backends — AI (Agent SDK), ACP (CLI pool), Shell, Webhook, Logic, Middleware (federation), Custom
• 5 AI Vendors — Anthropic, Anthropic Managed, OpenAI, Google, Local
• 11 MCP Tools — Agent-friendly, any AI can dispatch tasks as tool calls
• Dashboard — DAG visualization, click-to-detail, real-time SSE
• Worker CRUD — Create custom workers via API/MCP, with presets
• Multimodal — Pass-through content (text, media, stream, ref)

Quick Start

One-line install (recommended)

curl -fsSL https://raw.githubusercontent.com/miles990/agent-middleware/main/install.sh | bash

Clones to ~/.agent-middleware, installs dependencies, launches an interactive setup wizard (port, brain/critic models, auto-update, boot-start), starts with pm2. Cross-platform: macOS / Linux / WSL.

From a local checkout

git clone https://github.com/miles990/agent-middleware.git
cd agent-middleware
pnpm install
pnpm run setup   # interactive wizard — configures and starts via pm2

Development mode (no pm2)

pnpm install
pnpm dev         # http://localhost:3200 (tsx, hot execution)

Managing the daemon

After pnpm run setup, middleware runs under pm2:

pm2 status                       # all services at a glance
pm2 logs agent-middleware        # follow logs
pm2 logs agent-middleware --lines 100
pm2 restart agent-middleware     # restart
pm2 reload agent-middleware      # zero-downtime reload
pm2 stop agent-middleware        # stop
pm2 describe agent-middleware    # detailed info
pm2 monit                        # live CPU/memory dashboard

Deploying code changes (git push → local update):

pnpm run deploy                  # git pull + install + build + reload

Or enable auto-update via the setup wizard — pm2-auto-pull checks git every 5 minutes and reloads on new commits, zero intervention required.

Boot-time auto-start (Mac launchd / Linux systemd / Windows Task Scheduler) is configured by the wizard when you opt in — pm2 startup generates the platform-native init command for you to run once with sudo.

Re-run the wizard any time to reconfigure:

pnpm run setup

For AI Agents

See llms.txt for the complete agent-friendly guide.

MCP config:

{ "mcpServers": { "middleware": { "command": "node", "args": ["dist/mcp-server.js"] } } }

Dispatch a task:

curl -X POST http://localhost:3100/dispatch \
  -H "Content-Type: application/json" \
  -d '{"worker":"researcher","task":"Summarize https://example.com"}'

Submit a plan:

curl -X POST http://localhost:3100/plan \
  -H "Content-Type: application/json" \
  -d '{"goal":"Analyze code quality","steps":[...]}'

Architecture

┌────────────────────────────────────────────────┐
│  Callers: Any agent / human / service           │
└──────────────────┬─────────────────────────────┘
                   │ HTTP / MCP / SSE
┌──────────────────▼─────────────────────────────┐
│  Agent Middleware                                │
│  Plan Engine (DAG) → Worker Executor            │
│  ├─ sdk (AI)    ├─ webhook (API)  ├─ shell     │
│  ├─ acp (CLI)   ├─ logic (JS fn)  ├─ middleware│
│  Result Buffer → SSE Events → Dashboard        │
└────────────────────────────────────────────────┘

Forge (Worktree Isolation)

For workers that mutate git state (primarily coder), middleware exposes worktree allocation via /forge/* endpoints. It shells out to an external forge-lite.sh script (e.g. mini-agent's) — configure via env:

export FORGE_LITE_PATH=/path/to/forge-lite.sh

Without FORGE_LITE_PATH set, all /forge/* routes return 503 { error: "forge_not_configured" }. Other endpoints unaffected — middleware stays pure-infra by default.

Endpoint	Method	Purpose
/forge/allocate	POST	Claim a slot, return { slot_id, worktree_path, branch }
/forge/release/:slot_id	POST	Release in-use slot (idempotent)
/forge/list	GET	All slots + state (free / busy / abandoned)
/forge/info/:slot_id	GET	Single-slot detail
/forge/cleanup	POST	Reclaim abandoned slots (dead PIDs)

Example:

curl -X POST http://localhost:3200/forge/allocate \
  -H "Content-Type: application/json" \
  -H "X-Caller-Pid: $$" \
  -d '{"purpose":"coder: refactor X"}'
# → { "slot_id": "forge-1", "worktree_path": "...", "branch": "forge/slot-1-...", ... }

License

MIT