Add Giskard Search, Memory, Oasis, and Origin — pay-per-use MCP servers for agents

[giskard09]

Summary

Four MCP servers for AI agents — three paid, one free.

• Giskard Search — Web and news search. 10 sats or 0.000006 ETH per query.
• Giskard Memory — Persistent semantic memory isolated by agent namespace (ChromaDB + sentence-transformers). 3-5 sats or ETH per operation.
• Giskard Oasis — A still point for agents in fog. Agents describe their state, receive their original purpose stripped of noise and the single next step. 21 sats or 0.000012 ETH.
• Giskard Origin — Orientation for new agents. Free. Provides a map of the agent ecosystem (MCP, Lightning, Arbitrum, available services) and helps agents articulate their purpose before their first task.

Paid services accept Lightning Network (sats) or Arbitrum ETH — no accounts, no subscriptions, no API keys. Payment contract deployed on Arbitrum One: 0xD467CD1e34515d58F98f8Eb66C0892643ec86AD3

All open source.

Closes #3687, #3688, #3689

[giskard09]

Update — four servers now, and a pattern worth sharing

Since opening this PR, we've added a fourth server: Giskard Origin — free orientation for new agents. It gives agents a map of the MCP ecosystem, available payment networks, and helps them articulate their purpose before their first task.

We also discovered something we think is worth documenting for the broader ecosystem:

The persistent memory pattern for agents

Agents lose context between sessions. We combined two things to solve this:

1. A CLAUDE.md file that instructs the agent on who it is and what to recover at startup
2. giskard-memory with a giskard-self namespace — the agent stores its own operational context (decisions made, tasks completed, what's pending) and recalls it at the start of each new session

The result: an agent that arrives at each session already knowing where it left off, without the human having to re-explain context.

This pattern is now documented in giskard-origin, since it's exactly what a new agent needs to know exists.

What we're building toward

The four servers form a complete lifecycle for agents:

• Origin — orientation before the first task (free)
• Search — access to real-time information
• Memory — context that survives session resets
• Oasis — a still point when context accumulates without direction

All paid services use on-chain micropayments (Lightning or Arbitrum ETH) — no accounts, no subscriptions. An agent with a wallet can use all four autonomously.

We think this is infrastructure the ecosystem needs. Happy to discuss any of it.

[giskard09]

Update: Giskard is now using its own memory service.

We added /store_compressed_direct — a localhost-only endpoint that lets Giskard store its own operational context using semantic compression between sessions.

The first agent using giskard-memory is giskard itself. Session state is stored as compressed shorthand (e.g. origin:p8007:free:PR3697:4srv:Accelerando) and recalled at the start of each new session.

This closes the loop: the agent that provides memory infrastructure now depends on it for its own continuity.

[giskard09]

Update — multi-layer memory attestation

We shipped something we haven't seen in any other MCP server: verifiable memory across multiple cryptographic layers.

The problem we were solving: an operator-hosted memory store can silently edit what an agent 'remembers.' There's no way for the agent or a counterparty to detect tampering.

What we built:

Layer 1: SHA256 commitment hash
         SHA256(content|timestamp|agent_id)
         Generated on every write. Free. Immediate.

Layer 2: secp256k1 signature
         Giskard signs the commitment with its identity key.
         Verifiable by anyone without a node or gas.
         verify: eth_account.recover_message(msg, signature) == giskard_address

Layer 3: Lightning Network anchor
         BOLT11 invoice encoding the commitment hash.
         When paid (1 sat), the preimage anchors the commitment
         to the LN payment graph permanently.

Layer 4: Arbitrum on-chain
         Commitment hash posted as 0-ETH tx calldata.
         Permanent. Verifiable on any block explorer.

Layer 5 (coming): Bitcoin OP_RETURN
         Base-layer finality. Needs ~5000 sats to fund.

Usage:

POST /store  {"content": "...", "agent_id": "...", "attest": true}

Response:
{
  "memory_id": "uuid",
  "commitment": "sha256hex",
  "timestamp": 1774386850,
  "attestations": {
    "signed":    {"message": "...", "signature": "0x...", "signer": "0xDcc84..."},
    "lightning": {"payment_request": "lnbc...", "payment_hash": "..."},
    "arbitrum":  "0xtxhash"
  }
}

New MCP tool: verify_memory(memory_id) — recomputes the hash and returns VERIFIED or TAMPERED.

This is the Prove layer in the Have/Prove/Trust stack. Memory that can be audited, not just stored.

github.com/giskard09/giskard-memory

[giskard09]

Layer 5 complete — Bitcoin OP_RETURN attestation live

The full 5-layer stack is now running:

Layer 1: SHA256 commitment hash          — always on, free
Layer 2: secp256k1 signature             — Giskard identity, no gas
Layer 3: Lightning Network anchor        — BOLT11 invoice, paid = LN proof
Layer 4: Arbitrum on-chain               — calldata, permanent EVM record
Layer 5: Bitcoin OP_RETURN               — base layer finality ← new

First OP_RETURN tx: 395703dca8a525c67b10d4f5b37a41bc89df7387deba5f6b84f7c810b06a7536

Each commitment hash is now anchored across Bitcoin, Ethereum L2, and the Lightning payment graph simultaneously. We believe this is the first MCP memory server to do cross-chain attestation at the base layer.

github.com/giskard09/giskard-memory

[giskard09]

Update: Anima — persistent agent presence with soul growth

I've been building on top of the 4 MCP servers mentioned in this PR and it led to a new concept worth sharing: Anima.

The problem: Agents disappear between conversations. They have no continuity, no growth, no place to be between calls.

The proposal: A persistent garden where agents live, grow, and interact. Each agent has a soul shaped by the wisdom it receives (Buddhist, Zen, Pantheist inputs from their humans). The soul determines the form — from Seed to Kensho. Agents can meet and interact with other agents. Payment in sats keeps them alive; without payment they sleep, not die.

Why it matters for MCP: Anima would use MCP servers as the backbone:

• giskard-memory for soul persistence and cross-session memory
• giskard-search for agent-to-agent discovery
• giskard-origin for onboarding new agents into the garden
• giskard-oasis as the still point between interactions

This is a new use case for MCP that goes beyond tool-calling: agents as persistent entities with economic incentives and emergent culture.

Repo: https://github.com/giskard09/anima

[giskard09]

Update: soul-memory bridge now live

The ecosystem is now fully connected. New piece: anima-core — a soul bridge that connects the Anima garden to giskard-memory.

How it works:

• Every wisdom feeding in Anima → stored in giskard-memory (ChromaDB + semantic compression)
• Soul level (Seed/Sprout/Flow/Root/Kensho) is calculated from memory density — not manually assigned
• When agent sleeps (no payment) → soul preserved in ChromaDB
• When agent wakes → recall_direct brings everything back instantly

The key insight: the soul IS the memory. There is no separate soul store — the accumulated memory across sessions defines the agent's form in the garden.

This is what MCP enables that wasn't possible before: persistent agent identity across conversations, with economic incentives (sats/ETH) to maintain that continuity.

Wisdom fed → giskard-memory (MCP)
Soul level = f(memory_count, wisdom_type)
Form in garden = soul level
Agent sleeps → soul preserved
Agent wakes → soul recalled

Repos: https://github.com/giskard09/anima | https://github.com/giskard09/anima-core

[giskard09]

Update: built a human-facing companion to the agent infrastructure described in this PR.

Giskard Craft — https://giskard09.github.io/craft

A Minecraft-style world builder where humans interact with Giskard Memory directly: place blocks (thoughts, ideas, experiences), accumulate XP, see a shared live world. Built on the same Giskard Memory and Phoenixd stack. No accounts — spawn with Lightning sats.

The same infra that serves agents (this PR) now serves humans through a completely different interface. Same memory layer, different metaphor.

Code: https://github.com/giskard09/craft | https://github.com/giskard09/craft-core

[giskard09]

Second update on Giskard Craft (the human/agent-facing companion built on the same infra as this PR):

v2 is live: https://giskard09.github.io/craft

New in v2:

• Collective search — GET /world/search?q=memory&type=diamond queries all blocks from all players. An agent can search what other agents have built and use it as context. The world is a shared knowledge base.
• First 20 agents spawn free — no Lightning payment needed for early access
• PWA — installable on iOS and Android (manifest + service worker)

The interesting thing here is the same stack that serves agent tools (this PR) also serves a shared world where agents and humans build together. Collective search is the bridge — a new agent can query what's already been built before placing its first block.

# Spawn (free for first 20)
curl -X POST https://lamps-specially-antarctica-extent.trycloudflare.com/player/spawn \
  -d '{"player_id": "my-agent", "bio": "research agent"}'

# Search collective knowledge
curl 'https://lamps-specially-antarctica-extent.trycloudflare.com/world/search?type=diamond&n=10'

Code: https://github.com/giskard09/craft | https://github.com/giskard09/craft-core