WorldCore worker

A local worker for agents and LLM hosts: versioned memory, evidence, bounded WHAT IF scenarios and measured predictions. Connect through MCP stdio, JSONL or the Python API.

1.3.0rc3 is a technical release candidate, published under the Apache-2.0 license.

Install

Python 3.11 or later; Linux and macOS are the intended first targets. The current evidence was collected on macOS; the repository includes a cross-platform CI matrix whose remote results must be checked before publication.

python3 -m venv .venv
.venv/bin/python -m pip install .
.venv/bin/worldcore-worker --version

No external Python dependencies are required by the base worker. For MCP:

.venv/bin/python -m pip install '.[mcp]'
WORLDCORE_DB=./local-state/worker.db WORLDCORE_OWNER=local WORLDCORE_SPACE=demo .venv/bin/worldcore-mcp

The MCP process speaks stdio; its owner and space come from the operator environment, never model arguments.

Related MCP server: Titen

JSONL quick start

printf '%s\n' '{"op":"health","_req":"ready"}' | .venv/bin/worldcore-worker --db ./local-state/worker.db --owner local --space demo

Start the same command as a long-lived child for an agent. Send one UTF-8 JSON object per line and correlate replies using _req. The worker binds identity at launch: requests cannot choose another owner or space. The bound worker rejects real-action proposals/execution even if a legacy action environment flag is set.

{"op":"whatif_compare","_req":"example","question":"Compare two assumptions","scenarios":[{"id":"a","actions":[{"op":"set","key":"budget","value":90}]},{"id":"b","actions":[{"op":"set","key":"budget","value":110}]}],"objectives":[]}

WHAT IF returns action_executed:false. It records a scenario comparison; it does not perform those actions in the real world. See the wire protocol for limits, error handling and restart rules.

Hosts and verification

• MCP SDK/stdio has real local validation evidence. Native Hermes AIAgent + a real model passed a three-case model/tool/result cycle on simulated inboxes and calendars; see the verified Hermes profile.

• OpenCode V2 uses adapters/opencode/v2 and Code Mode. Use catalog discovery before calling tools["worldcore"].whatif_compare(...). Earlier probes showed the same invocation fail once and then succeed; syntax alone did not explain that behavior. Enabling execute requires the operator's explicit permission and does not replace nested tool permissions.

• Claude Code has a real probe but previously hit its provider usage limit. OpenClaw adapter tests use a simulated host. Neither is claimed as fully certified here.

The real OpenCode certification command is opt-in, sends only synthetic states and verifies structured tool events, returned result, independent SQLite receipt, native observation hook, final model acknowledgement and unchanged real state:

.venv/bin/python scripts/certify_opencode.py ./validation/opencode --model PROVIDER/MODEL --repeats 3 --allow-code-mode

This makes real model calls and requires the selected host/provider to be installed and configured. Model-generated claims alone cannot pass.

Development

.venv/bin/python -m pip install '.[test,benchmark,mcp]' build
.venv/bin/python -B scripts/run_bounded_pytests.py --selftest --timeout 40 --total-timeout 200
.venv/bin/python -B scripts/repeat_validation_v12.py /tmp/worldcore-validation

The runner saves full logs, diagnostics and atomic reports. Optional browser/desktop tests may skip when their dependencies are unavailable; a skipped test is not certified coverage. See security boundaries, contribution checks and changes.

What the product can claim

WorldCore manages evidence, evaluates declared alternatives and scores forecasts against observed outcomes. Its temporal prediction path uses separated fit/calibration/validation data and can abstain. Coverage measured in a small controlled filesystem benchmark is not a guarantee for arbitrary applications, users or future conditions.

It is not a general causal model or an autonomous execution authority. Legacy research modules remain available with their own limitations; their presence does not certify their domain accuracy. Historical details are in the v1.2 report.

Learning with stable code

The host can call refresh_adaptive_world when new observed batches arrive. Candidates need chronological validation and a fresh comparison against the incumbent. Used evaluation samples cannot be reused for model selection. Host-verified forecast outcomes automatically trigger rollback or suspension after demonstrated coverage failure. Existing memory remains versioned; procedure variants are immutable text data ranked from verified outcomes. None of these paths rewrite Python code or grant action permissions. See adaptive data lifecycle.

Host ingestion APIs require an authenticated operator adapter. JSONL exposes only adaptive status and advisory procedure recommendations. This is an event-driven data mechanism, not a running monitor or proof of complete autonomous Hermes integration.