Denario

Denario Assistant (Claude Code plugin)

Specialized Claude Code assistance for driving the Denario multi-agent research assistant over MCP — the end-to-end pipeline (idea → literature → methods → results → evaluate/iterate → paper → classify/publish) and the underlying cmbagent_lg analysis engine. This is the Claude-Code-native replacement for the OpenClaw gateway that fronts each Denario "scientist": the skills carry the same workflow knowledge and operating conventions, and the plugin bundles the MCP servers.

What you get

• /denario:explain — passive knowledge skill: the two MCP servers (denario, cmbagent_lg), the pipeline stages and which tool runs each, the params.yaml control surface (per-role/multi-provider models, max_n_steps, the VLM image-reviewer), and crash-recoverable restart. Auto-loads when the conversation is about Denario, denario_results/denario_idea/…, cmbagent_lg, params.yaml analysis models, or restart_at_step.
• /denario:conventions — passive operating-rules skill: stage discipline (one stage per turn, read-don't-assume), where logs/outputs land, the privacy + git/publish rules (private by default, confirm before going public), and the evaluate → iterate loop. The behavioral layer the autonomous scientist got from its SOUL.md/AGENTS.md.
• /denario:run-pipeline <project_dir> [data description] [--literature --vlm --iterate=K --no-citations --publish] — runs the pipeline end-to-end through the denario MCP tools, configured from params.yaml; reports a concise summary. Default run is idea → methods → results → paper (citations on by default); literature, iteration, publish, etc. are opt-in flags.

Prerequisite: the MCP servers (bundled)

The plugin auto-registers both MCP servers (denario, cmbagent_lg) via plugins/denario/.mcp.json — they start when the plugin is enabled. You only need to point them at your environment with two env vars (the servers ship with Denario, in denario/mcp_servers/):

# the python that has cmbagent_lg + Denario + modern langchain (the py312-cmbagent-lg venv)
export DENARIO_MCP_PYTHON=/Users/boris/pyvenvs/py312-cmbagent-lg/bin/python
# the directory containing denario_server.py / cmbagent_lg_server.py
export DENARIO_MCP_DIR=/Users/boris/GitHub/Denario/denario/mcp_servers

Export them where Claude Code can see them (your shell profile, or the project .env) before launching. The servers self-load GOOGLE_API_KEY from cmbagent_lg/.env; add OPENAI_API_KEY / ANTHROPIC_API_KEY there or export them too (stdio servers inherit Claude Code's environment).

Prefer to register them yourself instead? You can still do it manually and skip the env vars:

LGPY=$DENARIO_MCP_PYTHON; DEN=$DENARIO_MCP_DIR
claude mcp add denario     -s user -- $LGPY $DEN/denario_server.py
claude mcp add cmbagent_lg -s user -- $LGPY $DEN/cmbagent_lg_server.py

See /denario:explain → reference.md for keys, the venv, and gotchas.

Install (the plugin)

Local testing (no marketplace needed):

claude --plugin-dir ~/GitHub/denario-claude-plugin

From a marketplace (once published):

/plugin marketplace add borisbolliet/denario-claude-plugin
/plugin install denario@denario-claude-plugin

Try it

/denario:explain
How do I configure a multi-provider results run and turn on the image reviewer?

/denario:run-pipeline ~/Desktop/sho-study "A 1D damped harmonic oscillator: study amplitude/energy decay vs damping."

Layout

.claude-plugin/marketplace.json
plugins/denario/
  .claude-plugin/plugin.json
  .mcp.json             # bundles the denario + cmbagent_lg MCP servers (env-var configured)
  skills/
    explain/
      SKILL.md          # main knowledge skill (how to use the Denario MCP)
      reference.md      # tool signatures, full params.yaml schema, registration
    conventions/
      SKILL.md          # operating rules: privacy/git/publish, paths, iterate loop
    run-pipeline/
      SKILL.md          # end-to-end pipeline over MCP (+literature/iterate/classify/publish)

Notes

• MCP-first. Unlike the cobaya plugin (CLI via Bash), Denario is driven through MCP tools (mcp__denario__*, mcp__cmbagent_lg__*). The plugin bundles both servers in .mcp.json (configure with DENARIO_MCP_PYTHON / DENARIO_MCP_DIR); the run-pipeline skill pre-approves the pipeline tools in allowed-tools.
• denario_results is long and synchronous; for non-blocking runs use cmbagent_lg_execute + cmbagent_lg_status.
• Legacy stages. denario_eda and the cmbagent-keyword paper path need the old cmbagent package, which is not in the cmbagent-lg venv — they're out of scope here.

License

MIT