Redshank

An autonomous recursive language-model investigation agent written in Rust. Redshank ingests heterogeneous public datasets — campaign finance, lobbying disclosures, federal contracts, corporate registries, sanctions lists, court records, individual-person OSINT,and media intelligence — resolves entities across all of them, and surfaces non-obvious connections through evidence-backed analysis written into a live knowledge-graph wiki.

Redshank is a from-scratch Rust rewrite of OpenPlanter, replacing the Python runtime with a compiled binary that has zero Python or Node.js dependency.

Installation

cargo install redshank-cli --locked

Or build from source:

git clone https://github.com/greysquirr3l/redshank.git
cd redshank
cargo build --release

The binary lands at target/release/redshank.

Quickstart

# 1. Run the interactive credential wizard
redshank setup

# 2. Launch the TUI
redshank tui

# 3. Or run a one-shot investigation from the command line
redshank run "Who are the top donors to PACs linked to defense contractors with active SAM.gov registrations?"

redshank setup is the preferred entrypoint; redshank configure and redshank configure credentials are compatibility aliases for the same wizard. Secret fields are entered with hidden input, and unchanged values are detected so the command prints "No changes - nothing saved." instead of rewriting credentials.

Features

• Recursive tool-calling engine — The agent loop calls tools, reads results, and can delegate subtasks to child agent invocations with independent context windows. Context condensation keeps long investigations under the token limit.
• 90+ fetcher modules — Pull records from government databases, corporate registries, sanctions lists, court systems, and OSINT sources (see Data Sources below).
• Knowledge-graph wiki — Findings are written to interconnected Markdown documents with cross-references. A petgraph DAG tracks entities and relationships with fuzzy name matching.
• Interactive TUI — Three-pane ratatui interface: session sidebar, scrolling chat log, and a character-cell wiki-graph canvas. Slash commands for model switching, reasoning effort, and session management.
• Multi-provider LLM support — Anthropic (native Messages API with thinking budgets), OpenAI, OpenRouter, Cerebras, and Ollama (local).
• Security-first architecture — Fail-secure design with typed AuthContext, role-based SecurityPolicy, and chmod 600 credential storage. Every data access path checks permissions before touching storage.
• Stygian integration — Optional stygian feature flag enables stygian-graph pipelines and stygian-browser anti-detection automation for JS-rendered pages.
• Coraline MCP tools — Optional coraline feature flag adds code-aware file reading, semantic search, repo mapping, and file editing via the Coraline MCP server.
• CQRS + domain events — Every mutating operation flows through idempotent Command handlers; every read through Query handlers. State transitions emit typed domain events persisted to SQLite.

Data Sources

Configuration

Credentials are resolved in this priority order (first match wins):

1. REDSHANK_<KEY> — app-namespaced, useful when running multiple agents on one host
2. OPENPLANTER_<KEY> — legacy backward compatibility
3. <KEY> — bare env var (sufficient for most users)

Set keys as environment variables or in a .env file:

cp .env.example .env
# edit .env with your keys
chmod 600 .env

For persistent storage, copy the example credentials file:

mkdir -p .redshank
cp credentials.example.json .redshank/credentials.json
chmod 600 .redshank/credentials.json

Redshank merges from all sources in order: env vars → .env → <workspace>/.redshank/credentials.json → ~/.redshank/credentials.json.

Model defaults live in <workspace>/.redshank/settings.json:

{
  "default_model": "claude-sonnet-4-20250514",
  "default_reasoning_effort": "medium",
  "providers": {
    "OpenAiCompatible": {
      "enabled": true,
      "protocol": "openai_compatible",
      "deployment": "local",
      "base_url": "http://localhost:11434/v1",
      "default_model": "llama3.2"
    },
    "OpenAI": {
      "enabled": true,
      "protocol": "openai_compatible",
      "deployment": "local",
      "base_url": "http://localhost:1234/v1",
      "default_model": "qwen2.5-coder:latest"
    }
  }
}

Provider routing stays in settings.json, while actual secrets remain in .redshank/credentials.json. That lets you point OpenAI at a local OpenAI-compatible server, or route Ollama to a non-default host, without copying API keys into general settings.

Override at runtime with CLI flags:

redshank run --model gpt-4o --reasoning high "Investigate ..."

TUI Guide

Launch with redshank tui (or redshank tui --session <id> to resume).

Slash Commands

Headless Mode

Run without the TUI for scripting or CI:

redshank run --no-tui "Investigate ..."

CLI Reference

redshank [OPTIONS] <COMMAND>

Commands:
  run        Run an investigation with a given objective
  tui        Launch the interactive TUI
  fetch      Run a supported data fetcher directly
  session    List, resume, or delete sessions
  configure  Interactive credential setup
  setup      Alias for configure
  version    Print version and build info

Global Options:
  -w, --workspace <PATH>    Workspace directory [default: .]
  -m, --model <MODEL>       Override the default model
  -r, --reasoning <LEVEL>   Reasoning effort (off|low|medium|high)
      --no-tui              Run headless (no interactive UI)
      --max-depth <N>       Maximum recursion depth for subtasks
      --demo                Demo mode (use mock model)

redshank fetch currently supports uk_corporate_intelligence and its kebab-case alias uk-corporate-intelligence.

Development

Prerequisites

• Rust 1.94+ (stable)
• SQLite (bundled via rusqlite)

Build & Test

cargo build --workspace
cargo test --workspace
cargo clippy --workspace -- -D warnings

Optional Features

# Build with stygian web automation
cargo build --workspace --features redshank-fetchers/stygian

# Build with Coraline MCP code navigation tools
cargo build -p redshank-core --features coraline

Stygian Setup (JS-heavy sources)

Several fetchers — state Secretary of State portals, county property systems, and social media profiles — require a real browser to extract content. These route through stygian-mcp when it is running, and fail-soft (return empty results with a warning) when it is not.

1. Install stygian-mcp:

   cargo install stygian-mcp --locked
   

2. Start it on the default port:

   stygian-mcp --port 8787
   

3. Build redshank with the feature enabled and confirm the TUI footer shows stygian: ▲.

The health probe hits http://127.0.0.1:8787/health by default. Run stygian-mcp on that host and port. Configuration of the probe endpoint via settings.json is planned for a future release.

See the Stygian Fallback architecture doc for full setup, troubleshooting, and licensing-boundary rationale.

Workspace Layout

redshank/
├── redshank-core/       Core library: domain model, ports, engine, tools, persistence
│   └── src/
│       ├── domain/      Pure types, zero I/O deps
│       ├── ports/       Trait interfaces (inbound + outbound)
│       ├── application/ CQRS command/query handlers, engine service
│       └── adapters/    LLM providers, tools, SQLite, wiki filesystem
├── redshank-tui/        ratatui terminal interface
├── redshank-fetchers/   34 data-source fetcher libraries
├── redshank-cli/        clap CLI entry point
└── plan.toml            Build plan (26 tasks)

Comparison with OpenPlanter

License

See LICENSE for details.