deepparallel 0.4.1

DeepParallel - a multi-model agentic coding CLI with cross-model Guardian review, served via Crowe Logic.

DeepParallel

A multi-model agentic coding CLI with cross-model Guardian review. Many models check the work; you ship with a second opinion. Served via Crowe Logic.

Install

uv tool install deepparallel        # recommended
# or
pipx install deepparallel --python python3.13

Then run deepparallel (or the short alias dp).

Note: on systems with a broken Homebrew Python 3.14, bare pipx install may fail to find a usable interpreter - use uv tool install or pin --python.

From source

uv venv
uv pip install -p .venv -e .

Configure

Set the backend env vars (a .env file in the working directory is loaded automatically):

# default backend: azure
AZURE_CORE_ENDPOINT=https://<resource>.openai.azure.com
AZURE_CORE_API_KEY=<key>
# optional overrides
DEEPPARALLEL_API_VERSION=2024-08-01-preview

# or route through the Crowe Logic Foundry control plane:
DEEPPARALLEL_BACKEND=foundry
FOUNDRY_BASE_URL=https://<control-plane>
FOUNDRY_API_KEY=<token>

Use

deepparallel                 # interactive agent chat
deepparallel run "question"  # one-shot, pipe-friendly (answer on stdout)
deepparallel tools           # list the agent's tools
deepparallel info            # model + backend status
deepparallel doctor          # diagnose config + reachability
deepparallel run --no-tools "..."   # plain chat, no tools
deepparallel run --yes "..."        # auto-approve tool actions
deepparallel review <file|--diff>   # cross-model review as a CI gate (Pro)
deepparallel audit <file>           # supply-chain gate: catch hallucinated deps (Pro)
deepparallel research conduit       # latent-relay research demo (needs [research] extra)

Supply-chain gate

LLMs invent package names that don't exist — and attackers register those names ("slopsquatting"). deepparallel audit <file> extracts the dependencies a file introduces (imports, requirements.txt, package.json) and checks each against the real registry (PyPI / npm), flagging any the registry has never seen. Exit 0 clean, 2 if a likely-hallucinated package is found — so it gates a commit or PR. The Guardian also runs this automatically on every edit: a hallucinated import forces a confirmation even under --yes.

Fusion (stacking models for stronger output)

deepparallel can stack the answerer with a separate reasoning model. All modes compose hosted backends (no GPU/weight-merging), so they are API-call stacking.

deepparallel run --fuse reason "hard question"     # reasoner thinks, answerer answers
deepparallel run --fuse escalate "question"        # answer first; escalate to reasoner only if unsure
deepparallel run --deep "question"                 # heavy: many models in parallel + a judge (slow)
deepparallel run --dual "A,B" "question"           # compare two deployments side by side
deepparallel run --dual "A,B" --synth "question"   # ...and synthesize a merged answer

--fuse also works in interactive chat. Set a default with DEEPPARALLEL_FUSION=reason.

Fusion-native UX (what single-model agents cannot do)

• Guardian review - before an edit (write_file / edit_file / ast_replace_symbol) is applied, a second model reviews the diff and its verdict (safe / risky: ... / bug: ...) is shown in the confirm card. Advisory: you still approve. Toggle with DEEPPARALLEL_GUARDIAN; pick the reviewer with DEEPPARALLEL_GUARDIAN_DEPLOYMENT.
• Consensus + divergence - run --deep prints a consensus: chip from cross-model agreement (agreement, not correctness), and on low agreement lists the dissenting candidates.
• Live dial - in interactive chat, /fast, /fuse, /escalate switch the fusion mode mid-session (shown in the prompt); /deep runs the next prompt as a multi-model query.

Agent and tools

deepparallel is an agent: it can call tools to inspect and change your project.

• Read-only (run automatically): read_file, list_dir, glob, grep, ast_symbols, ast_show_symbol, web_fetch, web_search, analyze_image.
• Mutating / executing (require confirmation): write_file, edit_file, ast_replace_symbol, run_shell, run_code.

web_search needs DEEPPARALLEL_SEARCH_API_KEY; analyze_image works out of the box on a multimodal deployment (override with DEEPPARALLEL_VISION_DEPLOYMENT).

In interactive chat, mutating actions prompt for y/n. In run (non-interactive) they are denied unless you pass --yes. ast_* tools are multi-language via tree-sitter; run_code executes in a Docker sandbox when available, else a timeboxed local subprocess.

Configuration reference

Variable	Purpose	Default
DEEPPARALLEL_BACKEND	azure or foundry	azure
AZURE_CORE_ENDPOINT / AZURE_CORE_API_KEY	Azure transport	(required for azure)
DEEPPARALLEL_API_VERSION	Azure API version	2024-08-01-preview
FOUNDRY_BASE_URL / FOUNDRY_API_KEY	control-plane transport	(required for foundry)
DEEPPARALLEL_TEMPERATURE	default sampling temperature	0.4
DEEPPARALLEL_MAX_TOKENS	response cap (large enough to write whole files)	8192
DEEPPARALLEL_THINK	surface reasoning stream	0 (answer-only)
DEEPPARALLEL_TOOLS	enable agent tools	1 (on)
DEEPPARALLEL_AUTO_APPROVE	auto-approve mutating tools	0 (off)
DEEPPARALLEL_MAX_STEPS	max agent tool-call rounds	12
DEEPPARALLEL_SHELL_TIMEOUT	run_shell timeout (s)	120
DEEPPARALLEL_SANDBOX	auto / docker / subprocess for run_code	auto
DEEPPARALLEL_PLAIN	force plain (non-rich) output	0
DEEPPARALLEL_FUSION	default fusion: off / reason / escalate	off
DEEPPARALLEL_REASONER_DEPLOYMENT	reasoner model for fusion	DeepSeek-R1-0528
DEEPPARALLEL_PARALLEL_MODELS	comma-separated chains for --deep	(three defaults)
DEEPPARALLEL_JUDGE_DEPLOYMENT	judge/synthesizer model	the primary deployment
DEEPPARALLEL_SEARCH_API_KEY	enables web_search (Brave Search API)	(unset)
DEEPPARALLEL_SEARCH_URL	search API endpoint	Brave web search
DEEPPARALLEL_VISION_DEPLOYMENT	multimodal model for analyze_image	Llama-4-Scout
DEEPPARALLEL_GUARDIAN	second-model review of edits before apply	1 (on)
DEEPPARALLEL_GUARDIAN_DEPLOYMENT	the reviewer model	the reasoner

DeepParallel is served via Crowe Logic infrastructure. https://crowelogic.com