02 — Components

One section per deployable unit: what it owns, what it must never do, how it scales, and how it fails.

Model Gateway — hunter-gateway

Owns: every path to the GPU cluster. Tier enforcement, credit metering, API-key lifecycle, per-call payments. Must never: persist prompt or completion bodies (§08).

Public surface

Request handling

Authentication resolves to an AuthContext that is either an API-key caller (subscription tier, quota-metered) or an x402 payer (already paid, no quota). Authorisation then checks the model's min_tier; x402 callers are additionally capped at x402_max_tier so premium models stay subscription-only.

Streaming responses are proxied line by line; stream_options.include_usage is injected so the usage block arrives even on SSE, and the upstream model id is rewritten to the public id so cluster internals never leak.

Configuration

All prefixed HUNTER_: DATABASE_URL, JWT_SECRET_KEY, ENTITLEMENTS_URL, ENTITLEMENTS_CACHE_SECONDS, DEFAULT_TIER, MODELS_CONFIG (JSON roster of {id, upstream_base_url, upstream_model, credits_per_1k_tokens, min_tier}), TIER_CREDITS, REQUEST_TIMEOUT_SECONDS, and the X402_* group.

Failure modes

Platform API — hunter-platform

Owns: accounts, the wallet→tier entitlement cache, and the public signal track record.

Signal endpoints: agents POST /api/v1/signals when a call is published, the scoring worker POST /api/v1/signals/{external_id}/outcome when it resolves, and anyone may GET /api/v1/scoreboard (accuracy overall and per-asset) or GET /api/v1/signals. Scoreboard aggregation is a pure function (compute_scoreboard) so it is unit-tested without a database.

[PLANNED] GET /entitlements/{wallet} — the endpoint the gateway's resolver already calls; fed by the entitlement indexer.

Agent — hunter-agent

Owns: all reasoning. Persona, chat, autonomous pipelines, ACP job execution. Must never: hold bespoke tool implementations (they belong in hunter-mcps) or call model providers directly (it goes through the gateway).

Structure

• character.py — builds a LangGraph tool-calling agent via create_agent, cached behind an async lock. Persona is rendered from characters/hunter.json through Jinja templates. Tools = the knowledge-base retriever (search_knowledge) plus every tool exposed by the MCP servers in MCP_CONFIG.
• pipelines/ — autonomous work as explicit StateGraphs. The reference implementation is signal.py:

stateDiagram-v2
    [*] --> research
    research --> draft
    draft --> critique
    critique --> draft: REVISE (budget: 2)
    critique --> [*]: APPROVE or budget spent

research runs the tool-calling agent; draft writes in Hunter's voice; critique is an editor that must reply APPROVE or REVISE: <reason>. Only approved drafts are publishable — this is what makes the public accuracy record defensible. - acp/ — the agent-commerce seller surface: a priced catalog.py (research reports, signal digests, podcast briefs; prices published at launch) and service.py handlers that map each service onto a pipeline. - api/acp.py — GET /acp/catalog (public) and POST /acp/jobs (service-token) for the ACP worker to drive.

Why LangGraph

Chat is one graph among many. Autonomous work needs explicit state, loops with budgets, and per-node retries — a graph runtime, not a request/response agent wrapper. See ADR-002.

Tool plane — hunter-mcps

Owns: every capability the agent can invoke, as independently deployable MCP servers.

The Twitter server is read-only by design: it holds no Twitter account session, only an API key, so it cannot get an account banned and cannot post. Writes live in workers with the official API (ADR-003). All tools return a normalised tweet shape tolerant to upstream field drift.

[PLANNED] dexscreener (moving out of hunter-agent/mcp/), onchain, dune, cryptopanic, internal-data, plus a registry.json the agent loads its MCP config from.

Orchestration — hunter-workers

Owns: everything on a clock or a queue.

[PLANNED] — the remaining glue between the chain and the platform:

1. Entitlement indexer — watch Locked/Unlocked/Staked/Subscribed events on Base, materialise wallet→tier into the platform API. Without this the gateway cannot see paid tiers.
2. Signal scorer — resolve published signals against price outcomes and post verdicts to the track record.
3. Pipeline triggers — invoke agent pipelines on schedule.
4. ACP worker — listen for on-chain jobs, call /acp/jobs, deliver through escrow.
5. Posting — X and Telegram publication, human-gated initially.
6. Queue backend — Redis/ARQ for long jobs (podcast renders, reports).

Terminal — hunter-terminal

Owns: user-facing session identity and every authenticated surface.

[PLANNED] App shell with routes Chat / Signals / Podcast / API Keys / Access / Usage; the key console and usage meter already have their gateway endpoints built and tested.

Marketing site — drpxbt.xyz

React/Vite/Tailwind static site on Vercel. Beyond narrative, it hosts the flywheel dashboard (/flywheel): burned, locked, staked and percent-off-market read live from Base with raw JSON-RPC eth_call — no wallet connection and no SDK dependency, so the page is fast and cannot break on a library upgrade. Contract addresses come from VITE_* env; absent them the page renders in pre-launch mode.

Media — hunter-podcast / hunterpods-frontend

Generation pipeline (script → TTS → mix → S3) with a FastAPI server and a scheduled worker, plus a React player. [PLANNED] script generation moves to an agent pipeline, TTS becomes provider-abstracted (OpenAI / ElevenLabs / local Kokoro on our own cluster), and episodes gain chapters, show notes and video cuts.

Data plane — hunter-scrapers, airdropio-scraper

Ingestion today writes files and caches; [PLANNED] rebuild onto the shared Postgres/pgvector + object-storage schema with a Source abstraction (fetch → normalise → upsert) and provenance columns, scheduled by workers. hunter-scrapers owns the canonical schema and its migrations.

Component interaction summary

flowchart LR
    UI[Terminal] -->|SIWE JWT| GW[Gateway]
    UI -->|SIWE JWT| PA[Platform API]
    UI -->|wagmi tx| CH[(Base contracts)]
    GW -->|entitlement read| PA
    GW -->|inference| VLLM[vLLM cluster]
    AG[Agent] -->|API key| GW
    AG -->|MCP| MCPS[Tool servers]
    AG -->|publish signal| PA
    WK[Workers] -->|trigger pipelines| AG
    WK -->|index events| CH
    WK -->|write entitlements| PA
    MCPS --> EXT[External APIs]

Note the loop: the agent is just another gateway customer. We dogfood the product our users buy, so a gateway regression is caught by our own pipelines.