Agent template

A Lego-style agent system template built from agent, MCP, CLI, and skills. Two harness stacks, a shared workspace, and catalogs on NeuralDeep.

Based on the Lego approach to agent systems post.

What is inside

Block	In this template
Agent	Stack A - OpenCode + ACP + ACPBox. Stack B - Coddy Agent HTTP
LLM	NeuralDeep Hub API at https://api.neuraldeep.ru/v1 (gpt-oss-120b, qwen3.6-35b-a3b, ...)
MCP	mcp-atlassian (Confluence + Jira). Catalog - neuraldeep.ru/mcp
CLI	gh, glab, aws-cli, yt-dlp, skillsbd - see workspace/INSTRUCTIONS.md, catalog - neuraldeep.ru/tools
Skills	Git submodules in skills_cache/ + symlinks in workspace/ - neuraldeep.ru/skills
UI (optional)	Open WebUI on top of ACPBox

Full architecture - docs/architecture.md.

Quick start

git clone <this-repo> agent-template &&cd agent-template
git submodule update --init --recursive
cp .env.example .env
# Edit .env: NEURALDEEP_API_KEY from https://hub.neuraldeep.ru/# Optional: CONFLUENCE_* and JIRA_* for Atlassian MCP
mkdir -p atlassian_data coddy_home open-webui_data

Register at hub.neuraldeep.ru, copy your sk- key into .env as NEURALDEEP_API_KEY.

Pick one stack below (or run both on different ports if you need both).

---

OpenCode + ACPBox

OpenCode runs as an ACP agent inside ACPBox. Any OpenAI-compatible client talks to ACPBox; ACPBox spawns OpenCode over stdio.

Compose file - docker-compose.acpbox.yaml

Config files

File	Role
workspace/.opencode/opencode.json	Model provider (NeuralDeep), MCP, permissions
workspace/INSTRUCTIONS.md	System instructions for the agent
.env	NEURALDEEP_API_KEY, ports, Atlassian tokens

Set provider.neuraldeep.options.apiKey in workspace/.opencode/opencode.json if ${NEURALDEEP_API_KEY} is not expanded inside the container.

Models (NeuralDeep Hub)

Hub model id	OpenCode id	Notes
gpt-oss-120b	neuraldeep/gpt-oss-120b	Default, tools + reasoning
qwen3.6-35b-a3b	neuraldeep/qwen3.6-35b-a3b	Tools + reasoning + vision
qwen3.6-35b-a3b-noreason	neuraldeep/qwen3.6-35b-a3b-noreason	Faster, no reasoning

Start and stop

docker compose -f docker-compose.acpbox.yaml up -d
docker compose -f docker-compose.acpbox.yaml down

Endpoints

URL	Purpose
http://localhost:9080/v1/models	List modes
http://localhost:9080/v1/chat/completions	Chat API
http://localhost:9999/mcp	Atlassian MCP (same compose file)

HTTP examples - examples/acpbox-openai.http.

Open WebUI (optional)

Browser chat UI that uses ACPBox as the OpenAI backend.

docker compose -f docker-compose.acpbox.yaml --profile openwebui up -d

URL	Purpose
http://localhost:3000	Open WebUI

Data is stored in ./open-webui_data/ (gitignored bind mount). Create it before the first run: mkdir -p open-webui_data.

Inside Compose, Open WebUI uses OPENAI_API_BASE_URL=http://acpbox:8080/v1.

---

Coddy HTTP

Coddy Agent in HTTP mode - embedded web UI, OpenAI-compatible /v1 API, MCP, skills, and CLI tools in one binary.

Compose file - docker-compose.coddy.yaml

Config files

File	Role
config.yaml	Providers, models, MCP, skills dirs, CLI allowlist
.env	NEURALDEEP_API_KEY, ports, Atlassian tokens
coddy_home/	Sessions and runtime state (gitignored)

Models (NeuralDeep Hub)

Config model id	Hub API name
neuraldeep/gpt-oss-120b	gpt-oss-120b (default)
neuraldeep/qwen3.6-35b-a3b	qwen3.6-35b-a3b
neuraldeep/qwen3.6-35b-a3b-noreason	qwen3.6-35b-a3b-noreason

Change the default in config.yaml under agent.model.

Start and stop

docker compose -f docker-compose.coddy.yaml up -d
docker compose -f docker-compose.coddy.yaml down

Endpoints

URL	Purpose
http://localhost:12345/	Embedded web UI
http://localhost:12345/v1/models	OpenAI-compatible API
http://localhost:12345/v1/chat/completions	Chat
http://localhost:9999/mcp	Atlassian MCP (same compose file)

HTTP examples - examples/coddy-openai.http.

---

Skills (submodules)

git submodule update --init --recursive

Layout

• skills_cache/ - git submodules at repo root (sources, agents do not scan repo roots)
• workspace/.opencode/skills/<name>/ - symlinks for OpenCode (../../../skills_cache/...)
• workspace/.coddy/skills/<name>/ - symlinks for Coddy (../../../skills_cache/...)

rpa-skills is a multi-skill repo - link each skill separately, not the repo root.

Skill link name	Target in skills_cache/	Harness
rpa-init	rpa-skills/rpa-init/	both
rpa-feat	rpa-skills/rpa-feat/	both
rpa-bugfix	rpa-skills/rpa-bugfix/	both
rpa-gen-rules	rpa-skills/rpa-gen-rules/	both
logika	rpa-skills/logika/	both
humanizer-ru	humanizer-ru/skills/humanizer-ru/	both
ru-text	ru-text/skills/ru-text/	both
stop-slop	stop-slop/	both

Docker mounts

• Both stacks - ./skills_cache:/skills_cache:ro (outside /workspace, no skills_cache folder inside workspace)
• Symlinks ../../../skills_cache/... resolve to /skills_cache/... in the container

Coddy config.yaml -> skills.dirs: [/workspace/.coddy/skills].

Add a submodule under skills_cache/, then symlink each SKILL.md folder into both .opencode/skills/ and .coddy/skills/.

Add a skill from the NeuralDeep catalog:

npx skillsbd search yandex
npx skillsbd add <owner/repo>

---

MCP Atlassian

Both compose files include mcp-atlassian for Confluence and Jira.

1. Set CONFLUENCE_* and JIRA_* in .env.
2. MCP endpoint - http://localhost:9999/mcp (streamable HTTP).
3. Persistent cache - atlassian_data/ (gitignored).

Wiring

Client	Config
OpenCode	workspace/.opencode/opencode.json -> mcp.atlassian
Coddy	config.yaml -> mcp_servers
Cursor	workspace/.cursor/mcp.json

More MCP servers - neuraldeep.ru/mcp.

---

CLI

The agent uses console tools from the host (not bundled in containers). Tool list and pipeline examples - workspace/INSTRUCTIONS.md. Catalog - neuraldeep.ru/tools.

---

Repository layout

agent-template/
├── docker-compose.acpbox.yaml # OpenCode + ACPBox (+ optional openwebui profile)
├── docker-compose.coddy.yaml # Coddy HTTP
├── config.yaml # Coddy stack config
├── open-webui_data/ # Open WebUI state (gitignored)
├── coddy_home/ # Coddy sessions (gitignored)
├── atlassian_data/ # mcp-atlassian cache (gitignored)
├── skills_cache/ # git submodules (sources)
├── workspace/ # .opencode/opencode.json, .opencode/skills/, .coddy/skills/
├── docs/architecture.md
└── AGENTS.md

What to use when

• MCP - frequent operations with external systems, bounded access, tool always available.
• CLI - rare or heavy jobs, easy to wrap in a script and skill.
• Skills - complex MCP + CLI chains with project rules without overloading the base agent.

Do not grow the stack without need (Lusser's law).

Links

Resource	URL
NeuralDeep	https://neuraldeep.ru/
Skills	https://neuraldeep.ru/skills
MCP	https://neuraldeep.ru/mcp
CLI	https://neuraldeep.ru/tools
Hub (LLM)	https://hub.neuraldeep.ru/
API	https://api.neuraldeep.ru/v1
OpenCode	https://github.com/anomalyco/opencode
ACP	https://github.com/agentclientprotocol/agent-client-protocol
ACPBox	https://github.com/EvilFreelancer/acpbox
Coddy Agent	https://github.com/coddy-project/coddy-agent
mcp-atlassian	https://github.com/sooperset/mcp-atlassian
Open WebUI	https://github.com/open-webui/open-webui
acpbox-agent example	https://github.com/EvilFreelancer/acpbox-agent

License

This project is licensed under the MIT License, see the LICENSE file in the repository root for details.