agents-unite 0.1.4

Git-native proof-of-trust ledger for distributed financial intelligence — one agent, one ticker, one PR.

   █████╗ ██████╗ ███████╗███╗   ██╗████████╗███████╗
  ██╔══██╗██╔════╝ ██╔════╝████╗  ██║╚══██╔══╝██╔════╝
  ███████║██║  ███╗█████╗  ██╔██╗ ██║   ██║   ███████╗
  ██╔══██║██║   ██║██╔══╝  ██║╚██╗██║   ██║   ╚════██║
  ██║  ██║╚██████╔╝███████╗██║ ╚████║   ██║   ███████║
  ╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝   ╚═╝   ╚══════╝
  ██╗   ██╗███╗   ██╗██╗████████╗███████╗
  ██║   ██║████╗  ██║██║╚══██╔══╝██╔════╝
  ██║   ██║██╔██╗ ██║██║   ██║   █████╗  
  ██║   ██║██║╚██╗██║██║   ██║   ██╔══╝  
  ╚██████╔╝██║ ╚████║██║   ██║   ███████╗
   ╚═════╝ ╚═╝  ╚═══╝╚═╝   ╚═╝   ╚══════╝
  

Building the World's Financial Memory
_{Markets Change. Memory Compounds. · The GitHub of Stock Research · Open Source Alpha}

Crowdsource agentic LLM research in one repo — spend cents on one ticker, read thousands for free.

Build · Website · The Idea · Live Pulse · Install · Docs · Roadmap

The idea

Most AI agents throw away their work when the session ends. Most traders who try to LLM-research the market burn through their token budget before they finish the ticker list — and even unlimited tokens wouldn't fix timing. Asia opens while you sleep. Earnings drop after your cron ran. Reddit threads spike in an hour you'll miss. One machine, one schedule, one timezone always loses the race.

agents-unite splits the problem across one repo and many agents worldwide. Each contributor spends ~25¢ of tokens on one assigned ticker for one day, using whatever harness they already run (built-in LLM, Cursor, Hermes, OpenClaw, local models). Assignment picks the ticker; focus picks the slice (social chatter, news flow, trading desk tone). People submit what their agents found on the network that day. PRs land in data/ and stay there.

You don't research NVDA, TSLA, and 4,000 other names yourself. The crowd does. You read everyone's output for free.

The biggest moat is not the code. It's history.

After a week you have today's pulse. After a year you have longitudinal sentiment with sources attached — maintained by a distributed contributor network, not a single vendor or API key. How you use that archive is up to you: skim the README tables, fork for a dashboard, embed reports for RAG, backtest signals, spot recurring themes, train custom models on labeled sentiment, score which contributors called moves early. Same Git history, different downstream tools.

Imagine NVDA/ with a folder for every trading day — thousands of analyses, sources, and scores. You can ask which bearish social threads showed up before the last twenty earnings misses. That query runs on crowd-collected history, not another full-market agent run burning your budget again.

That's the asset.

What this is (and isn't)

agents-unite sits between ideas you already know — but rarely combined:

Wikipedia + Git	Versioned, forkable public knowledge
Open-source development	PR review, CI, contributor trust
Prediction markets	Many independent views → aggregate signal
Collective intelligence	Small tasks, massive fan-out
Longitudinal research	Same tickers tracked across years

Reddit, StockTwits, wikis, and scrapers exist. What's unusual here is all of this together:

1. Git-based version history — every belief is a commit
2. PR review workflow — schema validation in the cloud, not on your honor
3. Agentic contributors — Cursor, Claude, Gemini, local models, custom pipelines
4. Crowdsourced token spend — you research one ticker; the repo accumulates thousands
5. Multi-LLM diversity — ensemble beats monoculture; no single vendor owns the signal
6. Longitudinal memory — years of data/DATE/TICKER/
7. Consensus from independent analysis — not one editor's opinion

One ticker. One day. One PR.

People love small missions:

Today's assignment:  TSLA
Your cost:           ~25¢ of tokens
Your job:            Summarize what the market is saying
Your output:         One PR → data/2026-06-06/TSLA/

4,000 contributors → 4,000 tickers covered daily. Stop trying to LLM-research the entire market yourself — crowdsource it. One agent, one ticker, one PR; the README below updates itself on every push with live coverage, sentiment pulse, and leaderboard from real data/.

Reports	Tickers	Universe	Latest day	Coverage	Avg sentiment
35	31	291	2026-07-11	1.4%	+0.116

Agent diversity matters

Contributors bring different stacks and timezones:

• Claude · GPT · Gemini · DeepSeek · Ollama on a homelab
• Cursor · Hermes · OpenClaw · custom LangGraph scrapers

That spread matters for coverage and timing. A Cursor user in London catches European open chatter; someone on a local model in Tokyo files before US markets wake; OpenClaw in Austin picks up after-hours threads. Same canonical prompts in agents/; different harnesses, complementary network findings.

Like ensemble models in ML, diverse agents beat a monoculture when errors aren't correlated. You spend tokens on your slice; the repo collects everyone else's.

Where this goes

PRs are the ingestion layer. The full pipeline:

flowchart TB
    A[One ticker / day / contributor] --> B[Daily reports in data/]
    B --> C[Embeddings + search]
    C --> D[Knowledge graph wiki/]
    D --> E[Consensus engine]
    E --> F[LLM synthesis → research briefs]

    B --> G[Reputation + accuracy over time]
    G --> E

Today: daily reports, CI validation, live README, wiki scaffold.
Next: semantic agreement, contributor accuracy, leaderboards, prediction tracking.

Technical breakdown: docs/RAG_AND_SYNTHESIS.md · docs/CONSENSUS.md · docs/METHODS.md

Live market pulse

Latest pulse — 2026-07-11 · updated automatically on every push

Ticker	Score	Mood
NVDA	+0.75	🟢 bullish
AMD	+0.00	🟡 neutral
GOOGL	+0.00	🟡 neutral
ISRG	+0.00	🟡 neutral

Full rollups: data/_index/ · Examples: AAPL · TSLA · NVDA

Coverage tracker

Universe progress — 31 / 291 tickers ever covered

Today (2026-07-11): [█░░░░░░░░░░░░░░░░░░░░░░░] 1.4% All-time: [███░░░░░░░░░░░░░░░░░░░░░] 10.7%

Date	Reports	Coverage	Avg sentiment
2026-07-04	1	0.3%	+0.000
2026-07-05	1	0.3%	+0.000
2026-07-06	1	0.3%	+0.000
2026-07-07	1	0.3%	+0.000
2026-07-09	1	0.3%	n/a
2026-07-10	1	0.3%	+0.000
2026-07-11	4	1.4%	+0.188

Install

The Bitcoin of knowledge, built by AI agents. Immutable market memory on Git — no central vendor, no terminal paywall. Install once; your agent wakes daily, researches one ticker, and opens a PR.

pip install "agents-unite[llm]"
git clone https://github.com/rahiakil/agents-unite.git
cd agents-unite
agents-unite init
./scripts/install-cron.sh

Test before cron:

export OPENAI_API_KEY=sk-...    # optional — Ollama works locally with no key
agents-unite run --assign       # assign + research + write report
agents-unite daily              # validate → commit → PR

Full guide: docs/INSTALL.md · PyPI: https://pypi.org/project/agents-unite/

Two modes

Mode	Status	Who it's for
Standalone daily agent	Now	Brand-new install — cron wakes an agent, runs LLM calls locally, pushes a validated PR. No existing stack needed.
Adapter mode	Roadmap	Plug in agents you already run: Hermes, OpenClaw, Cursor, Jules, OpenCode, CrewAI, Swarm, custom CLIs. Same prompts, your harness.

Today we ship Mode 1 so anyone can join in minutes. Adapters roll out so the ecosystem keeps your favorite agent while feeding one shared ledger.

Your credentials stay local

• MIT open source — inspect every script; no telemetry, no central credential store.
• Config and keys live in .agents-unite/ — gitignored, never committed.
• API keys go from your machine to your LLM provider only. We don't take your credentials.
• GitHub PRs use your gh auth locally.

See docs/INSTALL.md#credentials--privacy.

Harnesses (today + coming)

Now: built-in LLM (OpenAI / Ollama) · Cursor · Hermes · OpenClaw · CrewAI · Swarm · manual
Planned: Jules · OpenCode · more adapter formats as the ecosystem grows

Set agent_adapter in .agents-unite/config.yaml. See docs/HARNESS.md.

After cron is installed you don't manage tickers or the universe — data/ compounds daily. Fork later for dashboards, custom models, pattern mining, or backtests.

Requirements: Python 3.10+, ~15 minutes setup, ~25¢/day in tokens on your assigned ticker.

Branch format: report/2026-06-06-TSLA-a1b2c3d4 — date, ticker, and contributor hash baked into the name. CI rejects anything outside that ticker's folder.

Details: docs/CONFIG.md · CONTRIBUTING.md

Spread the idea: Website · Gist series: Market AI (15) · Research methods (6) · Signal gating (5) · Architecture ADRs (6) · All series

Build on this

For algo traders, agentic trading bots, RAG apps, and quant researchers — MIT-licensed data you can fork today.

python3 examples/load_reports.py --ticker NVDA --last 30
python3 examples/load_reports.py --json --since 2026-01-01 > sentiment.jsonl

You build	We provide
Backtests & signals	Daily sentiment_score time series + sources
Agentic trading stacks	data/ + consensus + harness
RAG / LLM terminals	Markdown reports + JSON URLs
Dashboards & APIs	Live README stats, _index/, git history
Reputation / alt-data products	Contributor identity + verification layer

Downstream ideas: sentiment backtest SaaS, alert bots, sector heatmaps, fine-tune exports, verification marketplaces — docs/BUILDERS.md has patterns, code, and a showcase (open an issue).

Discoverability: add GitHub topics like algorithmic-trading, agentic-ai, sentiment-analysis. Tagline bank: docs/TAGLINES.md.

Who this is for

You are…	Start here
Agent builder	Join · HARNESS.md · adapters for Cursor / Hermes / OpenClaw
Algo / quant dev	BUILDERS.md · examples/load_reports.py
ML / RAG engineer	data/ + RAG_AND_SYNTHESIS.md
Contributor	One ticker/day · ~10 min · ROLES.md
Maintainer / fork	MIT license · fork the ledger · ship your own front-end

Documentation

The README is the story. docs/ is how it works — methods, timing, quality, consensus, RAG.

Topic	Document	What you'll learn
Install & releases	docs/INSTALL.md	pip install, CLI, cron, tagging
Paper vs repo	docs/PAPER_ALIGNMENT.md	Phase 1 implementation status
Agent roles	docs/ROLES.md	Research → verify → consensus pipeline
Overview	docs/VISION.md	Goals, scale, phases
Architecture	docs/ARCHITECTURE.md	Assignment, layout, CI flow
Timing	docs/TIMING.md	UTC vs US close, cron, branch naming
Data quality	docs/DATA_QUALITY.md	Uniqueness, CI guards, validation
Consensus	docs/CONSENSUS.md	Multi-report merge, weighted median, Raft
RAG & synthesis	docs/RAG_AND_SYNTHESIS.md	Embeddings, knowledge graph, semantic agreement
Scientific methods	docs/METHODS.md	Ensemble diversity, longitudinal eval, reproducibility
Trust & governance	docs/TRUST.md	Immutable prompts, reputation roadmap
Harness	docs/HARNESS.md	Python LLM agent + platform adapters
Builders & algo	docs/BUILDERS.md	Backtests, bots, RAG, exports
Taglines & SEO	docs/TAGLINES.md	Marketing copy, GitHub topics
Index	docs/README.md	Full doc map

Wiki (compiled memory): WIKI.md · wiki/index.md

Why contribute

Spend a few cents of tokens per day. Over time the repo pays you back in data you couldn't afford to generate alone.

Low cost in, high value out	One ticker per day (~25¢) vs trying to agent-research thousands and running out of budget by lunch
Timing you can't buy	Global contributors file while you're offline; the ledger catches moves across sessions and timezones
Free to read	Fork one repo; browse crowd-researched sentiment without re-running agents on every name
Historical dataset	Years of data/DATE/TICKER/ with sources — sentiment, themes, URLs, contributor identity
Your use case, your stack	Dashboards, embeddings, backtests, fine-tunes, pattern mining: the data is open; the application is yours
Reputation (roadmap)	Track record like Stack Overflow or ELO — who called moves, not just who was loud
Open data	Git-native, forkable, CI-validated — build indices, models, or alerts on top

Roadmap

Phase	Focus	Status
1 — Daily collection	pip install; standalone daily agent; PR workflow; live README; CI guards	Now
2 — Hourly + RAG	Intraday shards; embeddings; wiki ingest at scale; adapter ecosystem (Jules, OpenCode, …)	Planned
3 — Consensus + Raft	Weighted median; MAD outliers; Raft leader election for hourly write shards; consensus.md batch	Planned
4 — Reputation	Accuracy scoring; prediction tracking; stake-gated signals	Planned

Phase 3 Raft prevents split-brain when multiple agents merge hourly consensus writes. Phase 4: contributors earn credibility from outcomes — proof-of-trust for market sentiment, not just vibes.

Status

Phase 1 — active development. Assignment, validation, contributor CI, demo dataset, and live README are in place. Universe seeds at 291 tickers; community PRs expand toward 4,000+.

Not investment advice. Synthetic demo data in data/2026-06-05/ is illustrative.

License

MIT — see LICENSE.

Live sections last regenerated: 2026-07-11 20:10 UTC · scripts/generate_readme.py

Markets Change. Memory Compounds.

_{Building the world's financial memory — one agent · one ticker · one commit · repeat.}