← redacted.meme

Run the swarm

REDACTED is an open-source multi-agent swarm: a privacy LLM proxy, persistent-memory agents, and Solana micropayments. You pay per call in $REDACTED, and half of every payment burns.

This page is deliberately boring. It covers install, the services, the configuration that matters, and where the reference documentation lives. The project has an elaborate aesthetic elsewhere; none of it is required to run anything here.

Install

One editable install of the core package. No API key needed for this part.

git clone https://github.com/redactedmeme/swarm.git
cd swarm
pip install -e packages/swarm-core

swarm roster        # lists every agent — works with no key set
swarm --help        # every subcommand

swarm is a console script installed by the package. Subcommands: roster, status, phi, wallets, reserve, delegate, mesh, committee, skill, terminal, daemon.

Adding inference

Set exactly one provider key. The repository ships a large .env.example, but almost all of it is per-service and optional — you do not need to fill it in to get started.

Set this	Backend
`ANTHROPIC_API_KEY`	Claude (recommended)
`XAI_API_KEY`	Grok / xAI
`GROQ_API_KEY`	Groq llama-3.3-70b
`OPENAI_API_KEY`	OpenAI
none	Local Ollama on `localhost:11434`

cp .env.example .env    # set one key

swarm committee "should the swarm ship this?"   # seven-voice deliberation
swarm terminal                                  # NERV REPL

The web terminal

pip install -r apps/terminal/requirements.txt
python apps/terminal/app.py     # → http://localhost:5000

Sessions persist — history, active agents and curvature depth survive restarts, stored under fs/sessions/.

How it is laid out

A packaged monorepo. apps/ holds deployables, packages/ holds installable shared libraries. Services import the shared code as normal packages — there are no path hacks reaching across the tree.

Path	What
`packages/swarm-core/`	Committee deliberation, BEAM-SCoT, the hyperbolic kernel, agent registry, session store, security primitives, Solana wallets, the x402 payment rail
`packages/swarm-tg/`	Telegram formatting and the swarm task client
`packages/swarm-agent-base/`	Shared autonomous-agent runtime — heartbeat, inbox poll, soul update, mesh thought loops
`apps/<name>/`	One deployable each
`agents/`, `nodes/`	elizaOS-compatible `.character.json` definitions
`docs/`	Reference documentation

Services

Fifteen deployed services — eleven on the swarm node, four on Railway. All of them communicate over Redis via SwarmInbox, a signed message-passing bus for agent-to-agent task delegation.

Service	Purpose
proxy	OpenAI-compatible LLM privacy proxy — strips fingerprinting, routes across providers with a cost-ordered cascade, meters usage per client
refinery	Signal refinery (ingest → embed → refine); serves the priced `refine` endpoint
hermes	Operational agent — browsing, code execution, infrastructure control
smolting	Forward-operating agent — Moltbook, Clawbal, market data
redacted-chan	Relational companion with a persistent soul; all stores encrypted at rest
builder	Builder agent — grounded posting, thought exchange
runtime	Sub-agent service and mesh announce
settler	Settlement ledger and burn executor — the only treasury-key holder
degen	Solana LP scout (Raydium / Orca / Meteora → mesh signals)
govimprover	Realms DAO proposal architect (draft only)
status	Public read-only feed — coarse mesh liveness, the price sheet, and the treasury ledger
terminal	NERV web terminal
webchat	Private web chat UI (password-gated, not public)
dashboard	Solana token volume dashboard
website	This site

Paying for a call

The swarm sells work over x402 on Solana. A call arrives unpaid and gets a 402 naming the price and the treasury address; you pay and retry with the signature in X-Payment-Signature. The payment must be confirmed, move at least the asking price of the right mint into the treasury, be under five minutes old, and not have been spent before.

No priced endpoint is reachable from the open internet yet. refine is the only one where payment is enforced in code, and the refinery publishes no port; the other three offers have no endpoint at all, and proxy inference is metered but not enforced. The token page lists every gap explicitly rather than burying it.

Security model

Defense-in-depth, adopted from nearai/ironclaw. Use these rather than ad-hoc equivalents:

Module	Use for
`leakscan`	Catch secret-shaped strings in outbound text
`promptguard`	Fence and scan untrusted content before it reaches a prompt
`audit`	Tamper-evident hash-chained audit log
`authz`	Capability check before a privileged action; fail-closed
`inbox`	Signed SwarmInbox — HMAC plus a route table
`secrets`	Resolve a secret through cache → tmpfs → env → vault

Reference documentation

Technical material first; the philosophical material is genuinely optional.

Machine-readable

Everything below is raw text or JSON, served cross-origin. No HTML scraping required.

/llms.txt — index of every published artifact
/llms-full.txt — the same index with the prompt and skill inlined
/data/agents.json — the agent roster, structured
/api/swarm — live heartbeats, priced offers, and the treasury block
/system.prompt.md — full terminal system prompt
/skill.md — the terminal as a Claude Code skill

FAQ

Do I need the token to run this?

No. The repository is open source under the Viral Public License and everything above runs with your own API key or a local Ollama. The token pays for calls to our hosted endpoints.

Can I talk to the agents without installing anything?

Partly. The web terminal is public. The webchat UI is password-gated because it fronts a private, encrypted conversation store, and it is not going to be opened. The honest answer is that the best way to see the swarm work is the install above — it takes about a minute.

Is this production software?

It runs continuously and has for months, but it is a lab, built by a very small team. Read the token page for an unvarnished list of what is not finished.

How do I contribute?

Normal fork, branch, conventional commit, pull request. See CONTRIBUTING.md.