A guarded autonomous agent core for planning, execution, verification, maintenance, and bounded self-improvement.
ClawCore turns user goals into verified outcomes while staying bounded by fixed safety and approval policies. It is intentionally designed for strong autonomy inside clear guardrails, not unrestricted self-modification.
ClawCore is a bounded-autonomy runtime for structured task execution.
It is designed to:
- intake tasks in a structured format
- plan before long execution
- enforce approval and policy checks
- execute safe local work
- decompose larger tasks into child work
- verify outcomes before declaring completion
- learn from repeated failures inside approved boundaries
- Plan before long execution
- Verify outcomes before declaring completion
- Debug by root cause, not guesswork
- Learn from repeated failures and corrections
- Keep safety and approval policy fixed and auditable
- Allow limited self-improvement only in approved layers
- expand its own permissions
- rewrite safety policy autonomously
- self-replicate
- perform destructive or external actions without approval
- claim completion without verification
- structured task intake and schema validation
- simple web-app output is possible inside the ClawCore workspace (for example,
apps/paint/) - planner + orchestrator with goal signals, phase hints, replanning hints
- execution for command/file/directory/config/environment checks
- safe local fixes for JSON/YAML/Markdown/file/directory maintenance
- parent-child decomposition, generated subtask execution, and aggregation
- retry/backoff + blocked/permanent-blocked handling
- approval lifecycle: request / approve / reject / expire / rerequest
- approval audit trail and approval summary
- external-action mock execution with bounded external summary blocks
- lessons, candidate rules, adopted rules, maintenance triage, resolution planning
- dashboard summaries and structured report metadata (
schemaVersion: 0.2.0)
- Node.js with ESM support
- a local ClawCore workspace checkout
cd ClawCore
npm installnpm startnpm run run-allnpm run dashboardnpm run ui-operator -- --plan=operators/desktop/demo-open-safari.jsonFrom package.json:
npm start→node core/runner.jsnpm run dry-run→node core/runner.js --dry-runnpm run run-all→node core/run-all.jsnpm run approve-task→node core/approve-task.jsnpm run re-request-approval→node core/re-request-approval.jsnpm run adopt-rule→node core/adopt-rule.jsnpm run dashboard→node core/dashboard.jsnpm run migrate-approvals→node core/approval-migrate.jsnpm run ui-operator→node core/ui-operator.js
- load task
- classify + validate schema
- orchestrate + plan
- evaluate approval/policy
- execute task or child tasks
- verify outcomes
- review and optionally replan
- log lesson + write report
planner.jsexecutor.jsverifier.jsreviewer.jslearner.jsorchestrator.jsreplanner.jsapproval.jsmaintenance.jsdependency-view.jsexternal-report.jsreport-schema.js
core/— runtime modulespolicy/— fixed policy layerdocs/— architecture and implementation notesapps/— simple local app outputs and demosoperators/— desktop operator plans and examplestasks/queue/— queued tasks and fixture tasks tracked in the repo
The repository intentionally ignores local runtime state such as logs, memory, sandbox outputs, generated task artifacts, and report files.
ClawCore is now an operational bounded-autonomy core rather than only an MVP skeleton. Most real operating tasks are completing successfully; the remaining blocked tasks are intentional fixture cases for approval/schema handling and regression coverage.
- approval and external-action boundaries are intentionally strict
- some blocked tasks are preserved on purpose as regression fixtures
- desktop automation still depends on local OS permissions and environment setup
- this repository is a bounded local execution core, not an unrestricted autonomous agent