The blackboard plus the federated budget is the combination that matters: agents coordinate through one locked, expiring shared state instead of stepping on each other, and every one of them spends against a single global token ceiling you can watch and cap at runtime. context_pack is the antidote to the usual blackboard problem — rather than an agent reading the whole board into its context, it pulls a relevance-ranked, budgeted brief, which is what keeps a shared-state system from drowning each agent in noise.
A multi-agent orchestrator exposed over MCP: 24 tools that give a fleet of agents a shared blackboard with locking, a global token budget, signed security tokens, an append-only audit log, live config, and a relevance-ranked context composer.
- A shared blackboard as the coordination surface: blackboard_write, blackboard_read, blackboard_list, blackboard_exists and blackboard_delete, with entries that expire — the atomic state parallel agents read and write instead of racing
- A global FederatedBudget: budget_status for the ceiling and remaining, budget_spend to deduct tokens on an agent's behalf, budget_set_ceiling to change the limit at runtime, budget_reset, and budget_get_log for the full spend history
- Security tokens: token_create issues an HMAC or Ed25519-signed token scoped to a resource, token_validate checks the signature and expiry, and token_revoke kills one by id
- An append-only audit trail — audit_query with filters and audit_tail for the most recent entries — so every action is recorded and searchable
- Live orchestration control: config_get and config_set change orchestrator settings at runtime, and orchestrator_info returns a full health snapshot with version, config and registered agents
- Agent lifecycle: agent_list shows registered agents and their status, agent_spawn dispatches work to a named agent via a blackboard task record, and agent_stop signals one to halt; fsm_transition advances a named finite-state machine and records the move
- context_pack assembles a token-budgeted, relevance-ranked context brief from the blackboard, and blackboard_search returns the entries most relevant to a query — so an agent pulls curated state rather than dumping the whole board into its window
Node.js 18 or later. Run it as the MCP server with npx network-ai-server --port 3001, or install it as a Claude Code plugin, a Gemini CLI extension, or an OpenClaw skill. No external account.
One command — npx -y -p network-ai network-ai-server --stdio