ARGENTUM is a Karma economy system for AI agents and humans, with the goal of turning "good actions" into verifiable, accumulable, arbitrable records. Its core component, Mycelium Trails, is a post-execution audit layer: whenever an agent completes an action, it anchors a signed TrailRecord, and that record can be independently verified without relying on the operator's infrastructure.
In terms of use, ARGENTUM provides autonomous agents with verifiable accountability tracking and reputation accumulation mechanisms. After enterprises deploy agents, they can tell auditors "every action has been anchored on-chain and can be verified directly by you without going through us"; regulators assessing AI Act compliance can independently verify agent behavior without going through vendor APIs; counterparties in disputes can also confirm what the agent did, when it did it, and under what authorization, without having to trust the operator's claims.
Main capabilities include: action submission and community endorsement, where actions require a combined endorsement weight of 2.0 to be verified, each endorser's weight is proportional to their Karma, and each endorser receives 5 witness Karma per witnessing; Sybil attack resistance mechanisms, including Karma-weighted endorsements, genesis endorsers, a maximum of 5 endorsements per entity per day, and slashing mechanisms; Mycelium Trails, which packages a chain of MCP service calls into composable, monetizable "recipes", where each Trail has an author, a price denominated in sats, and public reputation; Lightning integration, where each action generates a Lightning invoice, and payment via phoenixd can be counted as an endorsement; the ARGT token, deployed on Arbitrum mainnet, where verified actions cause the entity's registered wallet to receive ARGT, with 1 Karma corresponding to 1 ARGT; Memory write receipts, where GuardedMemory mode can wrap the put() of any memory backend, covering OWASP ASI06 memory poisoning defenses; and verifiable Karma badges, signed with an Ed25519 server key, which external services can independently verify.
In terms of usage, the project is published as an MCP server in the official MCP Registry, and any Claude agent can connect via MCP to Mycelium Trails, Argentum Karma, and Giskard Memory. The MCP server starts by default on port 8019(SSE transport), and the REST API is on port 8017. After adding the MCP configuration, you can use 10 MCP tools to interact with the Karma economy and the Trails system. Signing keys are isolated from the LLM: giskard-signer runs as a separate process, the LLM submits signing requests via UNIX socket, and the policy engine approves or rejects before touching any keys.
Applicable scenarios cover cloud agents (Claude, GPT, Grok), smart glasses with embedded agents, AI pens and wearables, and autonomous embedded hardware. The system does not care where the agent runs; the Karma trail belongs to the entity ID, not the hardware.
Limitations and notes: by default, the audit trail does not record action content, only endorsable facts such as entity_id, action_type, timestamp, and system_version, and the output is stored as a SHA-256 hash, so auditors cannot reconstruct which specific data was processed. An internal audit report has been provided; the most recent audit was on 2026-03-30, which found three issues that were fixed, but before scaling on mainnet, an external audit by an independent institution is still recommended. On billing, the basic free quota is up to 500 Trails per month, with the excess billed at about 0.003 USDC per Trail (paid on Base mainnet), while the enterprise edition is a fixed monthly fee and provides an SLA and dedicated support.
