Christian Bucher

Vienna, AustriaOpen source systems

I build the trust layer around AI agents.

Memory with provenance. Outcomes observed independently. Claims backed by executable evidence. Behavior kept stable across model and prompt changes.

Explore Soul MCP Explore the full stack

AI-assisted. Architecture, verification, and release decisions owned by Christian Bucher.

Abstract Soul ledger network connecting memory, decisions, evidence, and outcomes

01 / Systems I build

Flagship / Public

Soul MCP 4.0.1

A trust-aware runtime for decisions, memory, skills, and verifiable outcomes. Local-first, inspectable, and built for any MCP-compatible AI client.

  • 23 MCP tools
  • 358 tests
  • MIT licensed
  • 0 required cloud services
claude mcp add soul -- npx -y soul-mcp
Source npm 3-minute start
MEMORY LEDGERHASH CHAIN
  1. 00017f3a9c1eDECISION
  2. 0002b2d4e7a9ACTION
  3. 0003c9f0b6d1EVIDENCE
  4. 0004e3a8d2f4OUTCOME

Integrity chain intact

Outcome layer / Public

Postcondition 0.1.0

Agents call tools. Postcondition checks whether the world changed. It declares a testable result, observes file, HTTP, Git, or npm state through constrained read-only verifiers, and preserves an honestly classified receipt.

  • MCP + SDK + CLI
  • 53 tests
  • Node 20 / 22 / 24
  • MIT
npx -y postcondition-mcp serve
Source npm Security boundary
POSTCONDITIONEXTERNAL OBSERVATION

“Version 0.1.0 is visible in the public npm registry.”

verifier
npm.version_exists
expected
0.1.0
observed
0.1.0
verdict
SATISFIED

receipt  e72168e33d15…

03 / Evidence contracts

Proofspec

Turns README, release, and product claims into versioned evidence contracts, constrained checks, receipts, and CI gates.

70 tests · 98.05%

04 / Compatibility gate

Behaviorlock

Deterministically compares recorded agent traces and keeps declared behavior stable across model, prompt, memory, policy, and tool changes.

73 tests · 9 matchers

05 / Operating contracts

Agent Invariants

Checks explicit approval, stop, scope, recovery, tool-use, and completion rules over normalized agent event traces.

36 tests · source release

06 / Stateful agents

ANIMA

A zero-dependency Python cognitive-state engine with durable state, associative memory, temporal context, model adapters, and ablation tooling.

446 tests · experimental

07 / Discovery layer

NexTool

A fast, searchable library of private browser utilities and practical technical guides—with dedicated launch pages for the trust stack.

269 tools · 131 guides

08 / Origin system

Miguel / Soul 4

The private personal-AI system from which the public memory, verification, evidence, and compatibility layers were extracted and hardened.

Private system · public primitives

Public when it is useful. Claimed when it is evidenced. Promoted when it is ready.

02 / How I build

Evidence before confidence.

I build systems that make decisions traceable, capture what happened, and improve with every verified outcome.

  1. 01

    Frame the decision

    Define the objective, constraints, and success criteria.

  2. 02

    Run the system

    Execute with tools, models, and data. Log the context.

  3. 03

    Verify the world state

    Observe the result independently of the action's own success response.

  4. 04

    Learn from the outcome

    Evaluate what worked, record why, and update the system.

DECISIONShip release 4.0.17f3a9c1e
ACTIONRun full release gateb2d4e7a9
EVIDENCEExternal state observedc9f0b6d1
OUTCOMERelease or repaire3a8d2f4

AI can accelerate the work. Ownership, testing, and release decisions stay human.

Soul preserves why. Postcondition checks whether the intended result exists. Proofspec binds public claims to evidence. Behaviorlock and Agent Invariants protect observable behavior. Each verified outcome becomes part of the next better decision.

03 / Open source & collaboration

Build with evidence.

Explore the systems, inspect every trust boundary, or add the published memory and outcome layers to an MCP client.

Install Soul MCP Install Postcondition GitHub npm Email

Next from the reservoir: confidence calibration, recovery contracts, model routing, and verified learning loops — promoted only when the evidence is ready.