SynKraken
howardscott-dot/synkraken
Plano de controle local para orquestrar e governar workforces de IA heterogêneos
Ver no GitHubSistema operacional de código aberto para controlar múltiplos agentes de IA heterogêneos (Claude Code, Goose, Hermes, OpenClaw, etc.) em um plano de controle local. Oferece visibilidade, governança, coordenação e recuperação de falhas através de interface terminal, web e CLI, sem substituir os runtimes existentes.
Instalação
npx convez-ai howardscott-dot/synkraken1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Coordenação de múltiplos agentes de IA em paralelo com aprovação humana
- Rastreamento de execução, auditoria e replay de operações de agentes
- Gerenciamento de saúde e recuperação de falhas de runtimes de IA
- Governança de propostas e decisões em workflows com múltiplos agentes
Veredito

SynKraken é uma solução robusta e bem-arquitetada para operadores que gerenciam múltiplos agentes de IA em produção. Vai além de simples orquestração: fornece rastreabilidade, governança e recuperação de falhas críticas para ambientes multi-agent complexos. Recomendado para times que precisam de controle operacional rigoroso.
README
A tradução para o português deste README ainda está na fila.
SynKraken
Open-source AI Workforce Operating System.
SynKraken is a local-first control plane for managing heterogeneous AI workers: Claude Code, Goose, Hermes, OpenClaw, Crush, Google Antigravity, and future CLI runtimes. It gives operators one place to coordinate workers, govern proposals, inspect runtime health, replay what happened, recover failures, and operate through terminal, web, CLI, and MCP-facing surfaces without replacing the runtimes they already use.
SynKraken is not a chatbot wrapper, an orchestration LLM, a CrewAI clone, a cloud SaaS product, or a hidden autonomous swarm. It is an operator system for AI work: workers may propose, humans approve, and SynKraken records the trail.
What SynKraken Is
SynKraken runs a local daemon that owns state in SQLite and exposes HTTP/SSE APIs to several operator surfaces: a terminal TUI, a local Web Command Deck, the CLI, and future MCP-compliant tools. Adapters invoke external AI runtimes through their native CLIs. The daemon records messages, deliveries, missions, outcomes, rooms, tasks, goals, decisions, handoffs, proposals, memory events, incidents, dead letters, and replayable traces.
Why It Exists
One-agent chat and single-agent IDE assistants are useful, but they are not enough for managing real AI work. Once several AI runtimes are involved, the operator needs durable rooms, visible assignments, approval boundaries, runtime reliability signals, failure recovery, and replayable evidence of what happened. Terminal multiplexers show shells. Workflow tools run steps. SynKraken manages the workforce state around AI work.
Feature Overview
SynKraken has three primary operator surfaces:
CLI → synkraken send, synkraken rooms, synkraken memory, etc.
Terminal TUI → Interactive dashboard with panels for workforce, rooms, events
Web Command Deck → Browser-based at http://localhost:9460 (when daemon is running)
MCP tools → Planned standards-compliant tool surface over daemon state
Key capabilities in brief:
- Workforce — discover and monitor registered AI runtimes
- Rooms — persistent, searchable collaboration rooms with transcripts
- Missions — governance containers for structured AI work
- Memory — operator-approved shared memory with bounded injection
- Flight Recorder — replayable trace of any message run
- Dead Letter Recovery — retry and replay failed deliveries
Core Capabilities
- AI workforce management across heterogeneous local runtimes
- runtime discovery and adapter-backed worker registration
- runtime reputation, trust scores, and workforce health
- mission governance containers for meaningful AI workforce outcomes
- outcome governance for desired results, evidence, confidence, and progress
- workforce assignments for visible ownership, contributors, blockers, review, completion, and handoffs
- persistent operational rooms with room memory and searchable transcripts
- bounded discussions, team tasks, and goal runs
- proposal approval governance for execution authority
- decision records and handoffs
- flight recorder replay, operational trace, and latest incident context
- dead letter capture, retry, and replay recovery controls
- peer-reviewed shared memory governance with bounded injection
- Shared Workforce Memory for visible, approved operator notes and scoped mission, outcome, assignment, room, runtime, and global context
- deterministic briefing, live activity, and operational awareness from daemon read models
- CLI, TUI, Web Command Deck, and future MCP-compliant surfaces
Architecture
Operator
│
├─ CLI
├─ TUI
├─ Web Command Deck
└─ MCP clients and tools
│ HTTP / SSE on loopback
▼
SynKraken Daemon
├─ HTTP API and event stream
├─ runtime fabric and dispatch
├─ proposal, governance, memory, trace, incident, and recovery models
└─ SQLite state store
│ subprocess adapter calls
▼
AI workers: Claude Code, Goose, Hermes, OpenClaw, Crush, Google Antigravity,
and future runtime adapters
The daemon is the source of truth. Operator surfaces render daemon state and invoke daemon APIs. SQLite is local, durable, and inspectable. Adapters stay small leaf integrations around external runtimes.
Quick Start
Install SynKraken with one command:
curl -fsSL https://raw.githubusercontent.com/howardscott-dot/synkraken/main/scripts/install.sh | bash
The installer creates ~/.synkraken, installs commands into ~/.local/bin,
auto-configures local supported workers, starts SynKraken, and prints the next
command to run.
From a development checkout:
./scripts/setup.sh
synkraken tui
If SynKraken is already installed and you want to refresh local worker setup:
synkraken setup
synkraken tui
For SSH workers or advanced rediscovery, use synkraken config.
See docs/INSTALLATION.md for Linux, macOS, uninstall,
diagnostics, and the Windows roadmap.
Basic operator commands:
synkraken send goose "summarize current status in one line"
synkraken send broadcast "status check"
synkraken memory note --title "Q2 engineering priorities" --body "Focus on API reliability and reducing deploy cycle time." --scope-type global --importance high
synkraken memory pending
synkraken memory approve <memory-id>
synkraken rooms
synkraken trace <id>
synkraken replay <id>
synkraken proposals
synkraken incident latest
synkraken retry dead-letter <id>
Configuring Adapter Paths
If your AI runtimes (Goose, Claude, Hermes, etc.) are installed in non-standard
locations, copy examples/config.paths.local.example.json to your working
directory and pass it when starting the daemon:
synkraken run --config config.paths.local.example.json
Edit the adapters.<name>.command array to use the full path to each binary,
for example ["/home/myuser/.local/bin/goose"] instead of ["goose"].
Retired Console Prototype
The native Tauri Console in apps/console is retired as an active product
surface. It remains in the repository as prototype/design history only. New
operator work should target the daemon API, CLI, TUI, Web Command Deck, and
future MCP-compliant tools. See
docs/CONSOLE_RETIREMENT.md.
Shared Workforce Memory is visible and governed. Approved memory can be
retrieved and injected into worker dispatches; proposed, rejected, and archived
records remain inspectable for audit. See
docs/WORKFORCE_MEMORY_MODEL.md.
The Activity screen is deterministic and daemon-backed. It shows recent runtime, room, governance, and failure events newest first, with runtime, room, mission, outcome, assignment, and event type filters. It is an awareness surface only; it does not schedule, plan, execute, or summarize with an AI model.
Missions
Mission is the primary organisational object for SynKraken v0.9. A mission is a governance container around meaningful work, not a task, project, ticket, schedule, or board. Missions link workers, rooms, traces, incidents, proposals, and relationships so the operator can answer what the workforce is doing and which outcomes are progressing.
The daemon exposes mission read models through /v1/missions,
/v1/missions/summary, and mission-scoped workers, activity, incidents, and
proposal endpoints. Console Mission Centre renders those records as an
operator cockpit beside Workforce, Rooms, Incidents, Proposals, and Traces.
Outcomes
Outcome is the primary success object for SynKraken v1.0. Workers perform activity, missions organise work, and outcomes measure whether the desired result was achieved. Outcomes are linked to missions and can connect to workers, traces, incidents, proposals, evidence, confidence, and status.
The daemon exposes outcome read models through /v1/outcomes,
/v1/outcomes/summary, mission-scoped outcomes, and outcome-scoped workers,
activity, incidents, and proposal endpoints. Console Outcome Centre is the
operator surface for seeing which results are completed, progressing, in
review, blocked, or waiting for approval.
Assignments
Assignment is the workforce accountability object for SynKraken v1.1. An assignment is work currently owned by one worker, with optional contributor workers and links to a mission, outcome, and room. It is not a task, ticket, schedule, board, or project plan.
The daemon exposes assignment read/write endpoints through /v1/assignments,
assignment-scoped activity, handoffs, proposals, traces, worker assignment
state, and mission/outcome/room assignment context. Console Assignment Centre
lets the operator create assignments, assign one owner, add contributors, mark
waiting or blocked, request review, complete work, and record explicit
handoffs. SynKraken does not automatically reassign, escalate, schedule, or
complete assignments.
Governance
SynKraken's execution authority model is:
Agents propose.
Humans approve.
SynKraken executes and records.
Workers can propose sensitive actions. Operators approve, reject, cancel, or execute through daemon governance endpoints. In v0.1, proposal execution is recorded as simulated execution; SynKraken does not grant workers autonomous shell, git, file, restart, retry, replay, or delete authority.
Flight Recorder And Trace
The Flight Recorder reconstructs work from existing durable records. Replay and trace views show messages, deliveries, dead letters, decisions, handoffs, tasks, goals, proposals, memory markers, runtime participation, failures, and outcomes. The purpose is operational understanding: what happened, which worker was involved, what failed, and what can be recovered.
Runtime Reputation
Runtime reputation is deterministic and delivery-derived. SynKraken tracks
successful replies, empty replies, timeouts, failures, wrong identity markers,
suspicious output, proposal outcomes, average duration, trust score, and health
status. Health statuses are healthy, degraded, unstable, and failing.
These signals make weak runtime behavior visible. They do not silently disable workers, hide failures, or replace operator judgment.
Incident Management
Failed deliveries and dead letters become inspectable operational facts. Operators can inspect latest incident context, open traces, retry failed deliveries, replay dead letters, and see how incidents relate to workers, rooms, proposals, and messages.
Operator Surfaces
SynKraken is daemon-first. The TUI is the primary terminal operator surface. The Web Command Deck is the browser operator surface. The CLI remains the scriptable surface. MCP compliance is the next standards-based integration surface. All of them call daemon APIs; none reads SQLite directly or owns a second state model.
Development
Runtime code is Python 3.10+ and stdlib-only.
Useful checks:
python3 scripts/smoke_test.py
python3 scripts/live_integration_test.py --skip-restart
python3 scripts/context_audit.py
python3 -m compileall synkraken scripts
The live integration test requires a running daemon. Individual subsystem smoke
tests live under scripts/.
Roadmap
Near-term direction:
- room-scoped delivery and runtime histories
- richer TUI and Web Command Deck workflows
- MCP-compliant tool surface over daemon state
- more adapter conformance coverage
- governed execution extensions beyond the current simulated proposal execution model, still under explicit operator authority
Running Multiple Instances
To run more than one SynKraken instance simultaneously, give each a unique
instance_name in the config:
{
"instance": {
"instance_name": "dev",
"organisation_name": "My Org",
"default_workspace": ""
}
}
Each instance gets its own SQLite database (synkraken-dev.db) and will not
conflict with other running instances. Attempting to start two instances with
the same instance_name will fail fast with a lock error.
synkraken run --config config-dev.json
synkraken run --config config-prod.json