parallect 0.2.0

Open-source multi-provider AI deep research CLI (BYOK)

parallect

Multi-provider AI deep research from the command line. Bring your own keys.

parallect fans a research query out to multiple frontier AI providers in parallel — Perplexity, Gemini, OpenAI, Grok, Anthropic, plus local models via Ollama and LM Studio — then synthesizes their outputs into a single report with cross-referenced citations, extracted claims, and consensus/conflict detection. The result is packaged as a portable, signed .prx archive you can read in any text editor, diff against another run, or publish to prxhub.

$ parallect research "What are the leading theories on dark matter?"

  Fanning out to: perplexity, gemini, openai, anthropic
  [████████████████] 4/4 providers  4m 32s  $0.18
  Synthesizing with anthropic ...
  Extracting 47 claims from 4 reports ...

  ✓ Saved to research.prx  (127 KB, signed)

  $ parallect research "..."  - dark matter theories
  → 4 providers  · 47 claims  · 23 sources  · 92% consensus

Why?

• No single provider knows everything. Different models see different sources, reason differently, and disagree in useful ways. Ask several at once, then see where they agree.
• Portable, signed output. Every run saves to an open .prx archive — plain text under the hood, verifiable via Ed25519, shareable via prxhub.
• BYOK. Your API keys, your bill, your data. No intermediate service sees your queries.
• Local-first option. Run entirely offline against Ollama or LM Studio if you don't want to call frontier APIs.
• Scriptable. Stable CLI, structured output, clean exit codes. Pipe it, chain it, call it from your editor.

Install

From PyPI:

pip install parallect

Or with uv:

uv tool install parallect

Requires Python 3.10+.

Quick start

# One-time setup — interactive TUI configures providers, keys, and defaults
parallect config

# Run research
parallect research "What are the leading theories on dark matter?"

# Pick specific providers
parallect research "quantum computing progress 2025" -p perplexity,openai,anthropic

# Deep research mode — premium models, higher cost, deeper reasoning
parallect research "CRISPR gene therapy developments" --deep

# Offline — Ollama or LM Studio only
parallect research "explain transformers" --local

# Follow-on research rooted in a prior bundle
parallect continue output.prx "What about practical applications?"

Commands

Command	Purpose
parallect config	Interactive TUI for providers, API keys, backends, plugins
parallect research <query>	Run a new research query
parallect continue <bundle> <query>	Follow-on research rooted in an existing bundle
parallect enhance <bundle>	Send a bundle to the hosted Parallect API for claim extraction + evidence graph
parallect jobs <subcommand>	Manage async SaaS-mode jobs (status, download)
parallect plugins <subcommand>	Manage data-source and pipeline plugins (list, status, index, config)

Run parallect <command> --help for full flags.

Providers

Provider	Default model	Deep model	Env var
Perplexity	sonar-deep-research	—	PARALLECT_PERPLEXITY_API_KEY
OpenAI	gpt-4o-mini	o3-deep-research	PARALLECT_OPENAI_API_KEY
Gemini	gemini-2.5-flash	gemini-2.5-pro	PARALLECT_GOOGLE_API_KEY
Anthropic	claude-sonnet-4	claude-opus-4	PARALLECT_ANTHROPIC_API_KEY
Grok	grok-3	grok-4	PARALLECT_XAI_API_KEY
Ollama	llama3.2	—	(local, no key)
LM Studio	default	—	(local, no key)
LDR	llama3.2	—	(local; pip install "parallect[ldr]")

Install optional provider SDKs with the all extra:

pip install "parallect[all]"

Configuration

parallect config writes to ~/.config/parallect/config.toml (or the platform equivalent). Precedence, highest to lowest:

1. Explicit CLI flags
2. Environment variables (PARALLECT_* prefix)
3. Project-local parallect.toml
4. User config
5. Defaults

Data source plugins

Plugins let parallect search and cite your local content alongside web sources. Built-in plugins:

Plugin	Purpose
filesystem	Index a directory of markdown / text / PDF
obsidian	Index an Obsidian vault, respecting links and graph ranking
prior_research	Reuse prior .prx bundles as a knowledge base
prxhub	Pull from a prxhub collection

Add one via parallect config → "Data source plugins" → "Add", or edit config.toml directly.

Pipeline plugins

External Python packages can hook into the research pipeline at four points:

class ResearchPlugin:
    async def pre_research(self, query: str, providers: list[str]) -> str: ...
    async def post_provider(self, provider: str, result: Any) -> Any: ...
    async def post_synthesis(self, synthesis: Any) -> Any: ...
    async def post_bundle(self, bundle: Any) -> Any: ...

Register via the parallect.plugins entry point. See docs/PLUGINS.md for the full protocol.

Output format

Every run saves to a .prx bundle — a gzipped tar archive:

research.prx
├── manifest.json          JSON-LD envelope: providers, cost, provenance
├── manifest.jws           Ed25519 signature of manifest hash
├── public-key.jwk         Public key for standalone verification
├── query.md               The research question
├── providers/
│   ├── perplexity/report.md + meta.json
│   ├── gemini/report.md + meta.json
│   └── ...
├── synthesis/
│   ├── report.md          Unified cross-provider synthesis
│   └── claims.json        Extracted atomic claims with provider attribution
├── sources/registry.json  Deduplicated, quality-scored sources
├── evidence/graph.json    Claim-to-source graph
└── provenance/graph.jsonld  W3C PROV-O provenance

Every file is human-readable text or JSON — tar -xzf research.prx and read in any editor.

Use the prx CLI to read, validate, diff, merge, sign, and publish bundles.

Documentation

• Quickstart — get running in 5 minutes
• Architecture — how the orchestrator works
• Providers — write a custom provider adapter
• Plugins — hook into the pipeline or add a data source

Development

git clone https://github.com/parallect/parallect.git
cd parallect
uv sync --group dev
uv run pytest tests/
uv run ruff check src/ tests/

Contributing

Contributions welcome — see CONTRIBUTING.md. For security reports, email [e-mail].

License

MIT — see LICENSE.

---

Built by SecureCoders. A hosted managed version is available at parallect.ai — same multi-provider pipeline, with billing, team workspaces, and a web dashboard.