Poggio AI

pAI MSc stores configuration in ~/.msc/config.yaml. You can manage it via the msc config CLI or by editing the file directly.

Quick Config via CLI

msc config get model
msc config set model claude-opus-4-6
msc config set budget_usd 50
msc config set output_format latex
msc config set counsel_enabled true

Environment Variables

API keys and service credentials are stored in ~/.msc/.env (created by msc setup). You can also set them as environment variables or in a project-level .env.

API Keys

Variable	Required	Description
`ANTHROPIC_API_KEY`	At least one	Claude models
`OPENAI_API_KEY`	At least one	GPT models
`GOOGLE_API_KEY`	At least one	Gemini models
`DEEPSEEK_API_KEY`	No	DeepSeek models

Optional

Variable	Description
`CONSORTIUM_SLURM_ENABLED`	Set to `1` to enable SLURM GPU job submission
`CONSORTIUM_TEXLIVE_BIN`	Path to TeX Live binaries
`LANGCHAIN_TRACING_V2`	Set to `true` to enable LangSmith tracing
`LANGCHAIN_API_KEY`	LangSmith API key
`SLACK_WEBHOOK_URL`	Slack notifications
`TELEGRAM_BOT_TOKEN`	Telegram notifications (with `TELEGRAM_CHAT_ID`)
`TINKER_API_KEY`	Tinker GPU API access

Never commit .env files.

LLM Config (`.llm_config.yaml`)

For advanced control, edit .llm_config.yaml in the project root. This overrides ~/.msc/config.yaml defaults.

main_agents:
  model: claude-sonnet-4-6
  reasoning_effort: high
  budget_tokens: 128000

budget:
  usd_limit: 25
  hard_stop: true
  fail_closed: true

Budget alerts trigger at 85%, 95%, and 100% of the limit.

Counsel Mode

counsel:
  enabled: false
  max_debate_rounds: 3
  synthesis_model: claude-sonnet-4-6

Per-Agent Model Tiers

per_agent_models:
  enabled: false
  tiers:
    opus: { model: claude-opus-4-6, reasoning_effort: high }
    sonnet: { model: claude-sonnet-4-6, reasoning_effort: high }
    economy: { model: claude-sonnet-4-6, reasoning_effort: low }

Experiment Tools

run_experiment_tool:
  code_model: claude-sonnet-4-6
  feedback_model: claude-sonnet-4-6
  vlm_model: claude-sonnet-4-6
  report_model: claude-sonnet-4-6

Cluster Config (`engaging_config.yaml`)

For SLURM/HPC deployments. See HPC Setup for full details.

cluster:
  name: engaging
  orchestrator:
    partition: your_partition
    time: "7-00:00:00"
    cpus: 4
    mem: "32G"
  experiment_gpu:
    partition: your_gpu_partition
    time: "7-00:00:00"
    cpus: 8
    mem: "64G"
    gres: "gpu:a100:1"

CLI Reference

Core Commands

msc setup
msc run
msc doctor
msc status
msc logs
msc runs
msc resume
msc campaign
msc config
msc budget
msc notify
msc openclaw
msc install

Use msc <command> --help for details.

Run Options

Flag	Default	Description
`--preset`	`standard`	`quick`, `standard`, `thorough`, `maximum`
`--task-file`	—	Path to a task file
`--output-format`	`latex`	`latex` or `markdown`
`--mode`	auto-detect	`local`, `tinker`, `hpc`

Stage Name Aliases

Alias	Full Stage Name
`literature`	`literature_review_agent`
`experiment`	`experimentation_agent`
`analysis`	`results_analysis_agent`
`writeup`	`writeup_agent`
`proofread`	`proofreading_agent`

Research Presets

Preset	Cost	Time	Best For
`quick`	$2-$5	~30 min	Testing, quick summaries, sanity checks
`standard`	$10-$25	~2 hrs	Most research questions, drafts
`thorough`	$40-$100	~6 hrs	Publication-quality drafts
`maximum`	$80-$200	12+ hrs	Rigorous manuscripts, comprehensive surveys

Supported Models

Provider	Models	Env Variable
Anthropic	`claude-opus-4-6`, `claude-sonnet-4-6`	`ANTHROPIC_API_KEY`
OpenAI	`gpt-5`, `gpt-5-mini`, `gpt-5.4`	`OPENAI_API_KEY`
Google	`gemini-3-pro-preview`	`GOOGLE_API_KEY`
DeepSeek	`deepseek-chat`	`DEEPSEEK_API_KEY`

Only one provider key is required. Multiple keys unlock counsel mode (multi-model debate).

Configuration