Seti

SETI

Default Workflow

1. Search: call web_search with a query
2. Filter: scan TOON results for relevant URLs
3. Enrich: call enrich_content on 1–10 URLs for full page text
4. Synthesize: use enriched content to answer or act

All output is TOON-formatted for minimal token usage.

MCP Tools

web_search

Search the web and return TOON-formatted results.

{ "query": "TypeScript best practices", "num_results": 20 }

Use wider sweeps (50+) for ambiguous or broad topics.

enrich_content

Fetch full page content for URLs via Jina Reader.

{ "urls": ["https://example.com/article"], "max_chars": 5000 }

Always enrich before citing specifics from a URL.

CLI Mode

seti "TypeScript best practices"
seti "rust vs go" 10
seti setup
seti verify

CLI output is also TOON-formatted.

Setup

seti setup                    # interactive
seti setup --non-interactive  # for agents
seti verify                   # check SearXNG health

SearXNG runs locally (Docker or uv). API keys are optional fallbacks.

Configuration

Optional API keys: TAVILY_API_KEY, EXA_API_KEY, BRAVE_API_KEY, GOOGLE_API_KEY, GOOGLE_CX, FIRECRAWL_API_KEY, SERPAPI_API_KEY, JINA_API_KEY.

Execution Rules

• Prefer web_search over browser automation for factual queries
• Use enrich_content when you need details beyond title + snippet
• Batch up to 10 URLs per enrich_content call
• Results are TOON by default — no parsing needed
• If search fails, check seti verify for backend health

Provider Fallback

SETI routes through SearXNG first, then falls back to configured API providers. Circuit breaker prevents hammering failing providers. Quota tracking persists at ~/.config/seti/usage.json.

Tech Stack

• Language: TypeScript
• Runtime: Bun
• Protocol: Model Context Protocol (MCP)
• Validation: Zod
• Lint/format: oxlint + oxfmt

Development Commands

bun install
bun run lint
bun run typecheck
bun run build
bun test
bun run smoke:mcp

Architecture

src/
├── mcp/          # MCP server and tool handlers
├── balancer/     # provider strategy, fallback, breaker, quota tracker
├── providers/    # provider-specific adapters
├── cache/        # in-memory cache
├── config/       # defaults + layered config loading
├── setup/        # setup and verify CLIs
└── utils/        # logging, validation, normalization helpers

Agent Guidelines

• Keep output format stable for MCP tools (web_search, enrich_content).
• Preserve fallback and quota semantics unless explicitly requested to change them.
• Validate provider-specific API parameter changes against upstream docs before implementation.
• Keep setup flows idempotent and non-destructive.
• Run lint, typecheck, build, and tests before finalizing changes.