turnstone 0.5.0

Multi-node AI orchestration platform with tool use, agent routing, and cluster simulation.

Turnstone

Multi-node AI orchestration platform. Deploy tool-using AI agents across a cluster of servers, driven by message queues or interactive interfaces.

Named after the Ruddy Turnstone — a bird that flips rocks to expose what's hiding underneath.

What it does

Turnstone gives LLMs tools — shell, files, search, web, planning — and orchestrates multi-turn conversations where the model investigates, acts, and reports. Native deferred tool loading for Anthropic and OpenAI APIs reduces token overhead and improves tool selection accuracy when MCP servers expose many tools; local models (vLLM, llama.cpp) get a transparent client-side BM25 fallback. It runs as:

• Interactive sessions — terminal CLI or browser UI with parallel workstreams
• Queue-driven agents — trigger workstreams via message queue, stream progress, approve or auto-approve tool use
• Multi-node clusters — generic work load-balances across nodes, directed work routes to a specific server
• Cluster dashboard — real-time view of all nodes and workstreams, workstream creation with node targeting, reverse proxy for server UIs (only the console port needs network access)
• Cluster simulator — test the stack at scale (up to 1000 nodes) without an LLM backend

Quickstart

Interactive (terminal)

pip install turnstone
turnstone --base-url http://localhost:8000/v1

Interactive (browser)

turnstone-server --port 8080 --base-url http://localhost:8000/v1

Queue-driven (programmatic)

pip install turnstone[mq]
turnstone-bridge --server-url http://localhost:8080 --redis-host localhost

from turnstone.mq import TurnstoneClient

with TurnstoneClient() as client:
    # Generic — any available node picks it up
    result = client.send_and_wait("Analyze the error logs", auto_approve=True)
    print(result.content)

    # Directed — must run on a specific server
    result = client.send_and_wait(
        "Check disk I/O on this server",
        target_node="server-12",
        auto_approve=True,
    )

Cluster dashboard

pip install turnstone[console]
turnstone-console --redis-host localhost --port 8090

Then open http://localhost:8090 for the cluster-wide dashboard. Create workstreams from the console and interact with any node's server UI through the built-in reverse proxy — no direct server port access required.

Docker

cp .env.example .env  # edit LLM_BASE_URL, OPENAI_API_KEY, etc.
docker compose up     # starts redis + server + bridge + console (SQLite)

For production with PostgreSQL:

# Requires POSTGRES_PASSWORD and DB_BACKEND=postgresql in .env (or exported)
docker compose --profile production up  # adds PostgreSQL, uses it as database

Console dashboard at http://localhost:8090. See docs/docker.md for configuration, scaling, and profiles.

Simulator

Test the multi-node stack at scale without an LLM backend:

docker compose --profile sim up redis console sim

Or standalone:

pip install turnstone[sim]
turnstone-sim --nodes 100 --scenario steady --duration 60 --mps 10

See docs/simulator.md for scenarios, CLI reference, and metrics.

All frontends connect to any OpenAI-compatible API (vLLM, NVIDIA NIM/NGC, llama.cpp, OpenAI, etc.) or Anthropic's native Messages API, and auto-detect the model.

Architecture

Diagrams

Detailed UML diagrams are available in docs/diagrams/:

Diagram	Description
System Context	Top-level components and external dependencies
Package Structure	Python modules and dependency graph
Core Engine Classes	SessionUI protocol, ChatSession, LLMProvider, WorkstreamManager
Conversation Turn	Full message lifecycle through the engine (provider-agnostic)
Tool Pipeline	Three-phase prepare/approve/execute
MQ Protocol	9 inbound + 19 outbound message types
Message Routing	Multi-node routing scenarios
Redis Key Schema	All Redis keys, types, and TTLs
Workstream States	State machine transitions
Simulator	SimCluster, dispatchers, scenarios
Console Data Flow	Dashboard data collection threads
Deployment	Docker Compose service topology
SDK Architecture	Python + TypeScript client libraries
Storage Architecture	Pluggable database backends (SQLite + PostgreSQL)

Multi-node routing

Each Turnstone server runs a bridge process. Bridges share a Redis instance for coordination:

Redis Key	Purpose
turnstone:inbound	Shared work queue — generic tasks, any node
turnstone:inbound:{node_id}	Per-node queue — directed tasks
turnstone:ws:{ws_id}	Workstream ownership — auto-routes follow-ups
turnstone:node:{node_id}	Node heartbeat + metadata for discovery
turnstone:events:{ws_id}	Per-workstream event pub/sub
turnstone:events:global	Global event pub/sub
turnstone:events:cluster	Cluster-wide state changes (for turnstone-console)

Routing rules:

1. Message has target_node → routes to that node's queue
2. Message has ws_id → looks up owner, routes to owning node
3. Neither → shared queue, next available bridge picks it up

Bridges BLPOP from their per-node queue (priority) then the shared queue. Directed work always takes precedence.

Tools

15 built-in tools, 2 agent tools, plus external tools via MCP:

Tool	Description	Auto-approved
bash	Execute shell commands
read_file	Read file contents	yes
write_file	Write/create files
edit_file	Fuzzy-match file editing
search	Search files by name/content	yes
math	Sandboxed Python evaluation
man	Read man pages	yes
web_fetch	Fetch URL content
web_search	Web search (provider-native or Tavily)
remember	Save persistent facts	yes
recall	Search memories and history	yes
forget	Remove a memory	yes
notify	Send notifications to linked channels	yes
task	Spawn autonomous sub-agent
plan	Explore codebase, write .plan.md
mcp__*	External tools from MCP servers

When the total tool count exceeds a configurable threshold (default 20), MCP tools are automatically deferred using native defer_loading on Anthropic and OpenAI APIs, or a transparent client-side BM25 search for local models. The LLM discovers deferred tools on demand via a tool_search capability — no configuration needed beyond --tool-search auto (the default).

MCP Tool Servers

Turnstone supports the Model Context Protocol (MCP) for connecting external tool servers. MCP tools are discovered at startup, converted to OpenAI function-calling format, and merged with built-in tools. Each MCP tool is prefixed with mcp__{server}__{tool} to avoid name collisions. Tool lists stay fresh via push notifications (tools.listChanged), periodic polling for servers without push, and manual /mcp refresh.

Configure via config.toml or --mcp-config:

[mcp.servers.github]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-github"]

[mcp.servers.github.env]
GITHUB_TOKEN = "ghp_..."

Or use a standard MCP JSON config file:

turnstone --mcp-config ~/.config/turnstone/mcp.json
turnstone-server --mcp-config ~/.config/turnstone/mcp.json

Use /mcp in the REPL to list connected tools, /mcp refresh to re-fetch tool lists from servers. MCP tools require user approval by default (overridden by --skip-permissions or UI auto-approve).

Multi-Model and Multi-Provider Support

Turnstone supports multiple model backends per server instance, including different LLM providers. ChatSession delegates all API communication to pluggable LLMProvider adapters — the internal message format stays OpenAI-like, and each provider translates at the API boundary. Define named models in config.toml and select per-workstream or switch mid-session with /model <alias>.

[models.local]
base_url = "http://localhost:8000/v1"
model = "qwen3-32b"
# provider defaults to "openai" (works with vLLM, llama.cpp, etc.)

[models.claude]
provider = "anthropic"
api_key = "sk-ant-..."
model = "claude-opus-4-6"
context_window = 200000

[models.openai]
base_url = "https://api.openai.com/v1"
api_key = "sk-..."
model = "gpt-5"
context_window = 400000

[model]
default = "local"              # which model to use by default
fallback = ["claude", "openai"]  # try these if the primary is unreachable
agent_model = "claude"         # optional: separate model for plan/task sub-agents

Supported providers: "openai" (default -- OpenAI, vLLM, llama.cpp, any OpenAI-compatible API) and "anthropic" (Anthropic Messages API, requires pip install turnstone[anthropic]).

Use /model to show available models, /model claude to switch. Workstreams created via the API accept an optional model parameter.

Configuration

All entry points read ~/.config/turnstone/config.toml. CLI flags override config values.

[api]
base_url = "http://localhost:8000/v1"
api_key = ""
tavily_key = ""        # only needed for local/vLLM models without native search

[model]
name = ""              # empty = auto-detect
temperature = 0.5
reasoning_effort = "medium"
default = "default"    # model alias for new workstreams
fallback = []          # ordered list of fallback model aliases
agent_model = ""       # model alias for plan/task sub-agents

[tools]
timeout = 30
skip_permissions = false
search = "auto"            # "auto" (enable when >threshold tools), "on", "off"
search_threshold = 20      # min tools before tool search activates
search_max_results = 5     # max tools returned per search query

[server]
host = "0.0.0.0"
port = 8080
max_workstreams = 10       # auto-evicts oldest idle when full

[redis]
host = "localhost"
port = 6379
password = ""

[bridge]
server_url = "http://localhost:8080"
node_id = ""           # empty = hostname_xxxx

[console]
host = "0.0.0.0"
port = 8090
url = "http://localhost:8090"  # used by CLI /cluster commands
poll_interval = 10

[health]
backend_probe_interval = 30
backend_probe_timeout = 5
circuit_breaker_threshold = 5
circuit_breaker_cooldown = 60

[ratelimit]
enabled = true
requests_per_second = 10.0
burst = 20

[database]
backend = "sqlite"     # "sqlite" (default) or "postgresql"
path = ".turnstone.db" # SQLite file path (relative to working directory)
# url = "postgresql+psycopg://user:pass@host:5432/turnstone"  # PostgreSQL
# pool_size = 5        # PostgreSQL connection pool size

[mcp]
config_path = ""       # path to MCP JSON config file (alternative to TOML sections)
refresh_interval = 14400  # periodic refresh for servers without push notifications (seconds, 0 to disable)

[mcp.servers.example]  # one section per MCP server
command = "npx"
args = ["-y", "@modelcontextprotocol/server-example"]
# type = "stdio"       # "stdio" (default) or "http"
# url = ""             # for HTTP transport

Precedence: CLI args > environment variables > config.toml > defaults.

Workstreams

Parallel independent conversations, each with its own session and state:

Symbol	State	Meaning
·	idle	Waiting for input
◌	thinking	Model is generating
▸	running	Tool execution in progress
◆	attention	Waiting for approval
✖	error	Something went wrong

Idle workstreams are automatically cleaned up after 2 hours (configurable). In multi-node deployments, workstream ownership is tracked in Redis — follow-up messages auto-route to the owning node.

Monitoring

/metrics endpoint exposes Prometheus-format metrics:

• turnstone_tokens_total{direction} — prompt/completion token counters
• turnstone_tool_calls_total{tool} — per-tool invocation counts
• turnstone_workstream_context_ratio{ws_id} — per-workstream context utilization
• turnstone_http_request_duration_seconds — request latency histogram
• turnstone_workstreams_by_state{state} — workstream state gauges
• turnstone_sse_connections_active — current open SSE connections
• turnstone_ratelimit_rejected_total — requests rejected by rate limiter
• turnstone_backend_up — LLM backend reachability (0/1)
• turnstone_circuit_state — circuit breaker state (0=closed, 1=open, 2=half_open)
• turnstone_workstreams_evicted_total — workstreams auto-evicted at capacity

Per-workstream metrics are labeled by ws_id (bounded to 10 max workstreams).

Health & Rate Limiting

Health degradation. A background BackendHealthMonitor probes the LLM backend every backend_probe_interval seconds. When the backend is unreachable, /health reports "status": "degraded" (HTTP 200) and the turnstone_backend_up gauge drops to 0.

Circuit breaker. After circuit_breaker_threshold consecutive probe failures the circuit opens (CLOSED -> OPEN). While open, ChatSession._create_stream_with_retry skips the backend entirely and returns an error. After circuit_breaker_cooldown seconds the circuit enters HALF_OPEN, allowing a single probe. A successful probe closes the circuit; a failure re-opens it.

Per-IP rate limiting. When [ratelimit].enabled is true, each client IP is tracked with a token-bucket limiter (requests_per_second / burst). Rate limiting is applied in do_GET/do_POST after authentication but before route dispatch. /health and /metrics are exempt. Requests that exceed the limit receive HTTP 429 with a Retry-After header.

Workstream eviction. When WorkstreamManager.create() would exceed max_workstreams, the oldest IDLE workstream is automatically evicted and the turnstone_workstreams_evicted_total counter is incremented. Configure via [server].max_workstreams (default 10).

Requirements

• Python 3.11+
• An OpenAI-compatible API endpoint (vLLM, NVIDIA NIM, llama.cpp, etc.) or an Anthropic API key
• Redis (for message queue bridge — pip install turnstone[mq])
• Anthropic provider (optional — pip install turnstone[anthropic])
• PostgreSQL (optional, for production — pip install turnstone[postgres])
• Git LFS (for cloning — diagram PNGs are stored in LFS)

License

Business Source License 1.1 — free for all use except hosting as a managed service. Converts to Apache 2.0 on 2030-03-01.