← Back to Skills Marketplace
Network-AI
by
Jovan Marinovic
· GitHub ↗
· v5.16.0
· MIT-0
7321
Downloads
6
Stars
0
Active Installs
165
Versions
Install in OpenClaw
/install network-ai
Description
Local Python orchestration skill: multi-agent workflows via shared blackboard file, permission gating, token budget scripts, and persistent project context. The bundled Python scripts make no network calls and have zero third-party dependencies. The parent repository also contains a TypeScript engine (not included in this skill bundle).
Usage Guidance
Install only in workspaces where local persistent coordination files are acceptable. Do not put secrets, credentials, or personal data in permission justifications or project context, and do not treat the advisory grant tokens as real authorization for payments, databases, email, or exports without separate human approval and authenticated identity checks.
Capability Assessment
Credentials
Filesystem access is local and disclosed, with data rooted under the skill project data directory or data/<env>; NETWORK_AI_ENV is validated and used only for state partitioning.
Install Mechanism
No install step or third-party package is required; requirements.txt is documentation-only and the scripts use Python standard library imports.
Instruction Scope
The skill gives strong orchestration workflow instructions and persistent-context injection guidance, but it scopes them to multi-agent coordination and includes warnings that stored context is untrusted reference data.
Persistence & Privilege
The skill persists blackboard entries, audit logs, advisory grants, a signing key, budgets, health state, and project context locally; this is expected for the skill but users should treat the stored context and audit text as sensitive local data.
Purpose & Capability
The capability set fits the stated purpose: local blackboard coordination, budget tracking, advisory permission scoring, token validation/revocation, and persistent project context for multi-agent workflows.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install network-ai - After installation, invoke the skill by name or use
/network-ai - Provide required inputs per the skill's parameter spec and get structured output
Version History
v5.16.0
5.16.0: opt-in ASI Strict Mode in the npm package (ASI_STRICT_MODE=audit or enforce; off by default, no behaviour change). Python skill scripts unchanged.
v5.15.6
ClawHub 5.15.5 scan follow-up: removed contradictory filelock advice (locking is built in), corrected script count and token-expiry wording, declared deletions and archival in filesystem_operations, scoped check_permission.py docstrings.
v5.15.5
ClawHub 5.15.4 scan follow-up: inject preamble reworded, invocation capability replaces autonomous_actions, blackboard.py archive writes declared, scan-findings history moved to SECURITY.md. context_manager.py CodeQL py/mixed-returns fix.
v5.15.4
Security: context_manager.py validates writes and scans every field for injection and role-delimiter patterns (A.I.G T02); inject output is delimited reference data. check_permission.py help aligned with HIGH_RISK_RESOURCES. SKILL.md declares allowed-tools and bundle-only capabilities, removes unpinned npx.
v5.15.3
v5.15.3: per-instance orchestrator token, MCP server-held identity, shell-free git provenance
v5.15.2
v5.15.2: security fixes GHSA-9p2w-prp8-5722, GHSA-hr6v-mfxm-4438, GHSA-4pvg-m42h-c3x2, CodeQL 179/180
v5.15.1
v5.15.1 security patch: fix GHSA-743h-jr5x-mpcr (ClaudeHookBridge deny-pattern bypass via truncation before the security decision) and GHSA-9v4f-j8cv-fhxw (SandboxPolicy blocklist/approval-gate bypass via quote/whitespace matcher mismatch); 3638 tests across 41 suites
v5.15.0
v5.15.0 context signal-over-noise: ContextComposer (token-budgeted, relevance-ranked context assembly with recency decay and scope affinity) plus context_pack and blackboard_search MCP tools (24 tools total); 3603 tests across 40 suites
v5.14.0
v5.14.0 ecosystem expansion: Gemini, OpenAI Responses, and Claude Agent SDK adapters (32 total); Claude Code hooks gating via AuthGuardian; MCP elicitation approvals; A2A server mode; Gemini CLI extension; AGENTS.md; MCP Registry manifest
v5.13.4
Security fixes: GHSA-m4jg-6w3q-gm86 (ApprovalInbox auth+CORS) and GHSA-3jf7-33vc-hgf4 (APSAdapter fail-closed signature verification)
v5.13.3
fix: correct display name to Network-AI
v5.13.2
fix: correct ClawHub display name
v5.13.1
fix: reword audit-log privacy note to clear SkillSpector YARA false positive
v5.13.0
v5.13.0: model-interaction lifecycle governance — GovernedModelGateway (refusal->fallback->billing), ModelBudget fallback-credit repricing, RefusalTelemetry, EffortPolicy, per-sub-agent fallback, ThinkingBlockManager, OWASP Agentic Top 10 coverage matrix
v5.12.7
v5.12.7: fix recurring SkillSpector finding — comment.txt draft note no longer bundled (ClawHub honours .clawhubignore); added npm run clawhub:check bundle-hygiene guard; closed data/, .env, *.log and other bundle leaks
v5.12.6
v5.12.6: CodeQL security fixes — resolve #177 (indirect command injection in socket-check.js: spawnSync + semver validation), #176 and #175 (unused imports). New scripts/codeql-check.js monitor + npm run codeql:check. SKILL.md: 3 new SkillSpector by-design findings documented. RELEASING.md: CodeQL gate step added.
v5.12.5
v5.12.5: Supply-chain security hardening. Remove gptSecurity (debugAccess) alerts — replaced String.fromCharCode obfuscation pattern in blackboard-validator with named EVAL_FN constant. Explicit policy gate at shell exec call sites in bin/console.ts. SUPPLY_CHAIN.md sections 5a (shell surface) and 5b (telemetry). New scripts/socket-check.js score monitor. Supply Chain Score: 75 → ~80.
v5.12.4
5.12.4 - SkillSpector findings resolved (.clawignore excludes comment.txt, SKILL.md trigger hardening), Socket.dev scan gap closed (9 missing entries: d.ts decls, ESM adapter mirrors, example/run shellAccess). networkAccess 64, shellAccess 10.
v5.12.3
5.12.3 - Socket.dev supply-chain triage for the dual CJS+ESM build (gptSecurity console pipe-mode entry plus dist/esm networkAccess and shellAccess mirrors) and console pipe-mode approval hardening (fail-closed: approval-required operations are denied unless --auto-approve).
v5.12.2
5.12.2: Security patch - fix 5 CVEs: symlink escape in backup (GHSA-6x2m), restore() path traversal (GHSA-48x2), poisoned manifest pruning (GHSA-2fmp), sandbox path-prefix bypass (GHSA-jvcm), unauthenticated ApprovalInbox (GHSA-mxjx).
Metadata
Frequently Asked Questions
What is Network-AI?
Local Python orchestration skill: multi-agent workflows via shared blackboard file, permission gating, token budget scripts, and persistent project context. The bundled Python scripts make no network calls and have zero third-party dependencies. The parent repository also contains a TypeScript engine (not included in this skill bundle). It is an AI Agent Skill for Claude Code / OpenClaw, with 7321 downloads so far.
How do I install Network-AI?
Run "/install network-ai" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Network-AI free?
Yes, Network-AI is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does Network-AI support?
Network-AI is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Network-AI?
It is built and maintained by Jovan Marinovic (@jovancoding); the current version is v5.16.0.
More Skills