Configuration Reference

The following is a complete reference of the plano_config.yml that controls the behavior of a single instance of the Plano gateway. This where you enable capabilities like routing to upstream LLm providers, defining prompt_targets where prompts get routed to, apply guardrails, and enable critical agent observability features.

Model provider headers

Each entry under model_providers (or the legacy llm_providers alias) may include a headers map of extra HTTP headers that Plano adds to upstream LLM requests. Plano applies these headers after it sets authentication from access_key or passthrough_auth, so you can supply provider-specific metadata without replacing the configured credentials.

  • Type: map of strings (header name → value)

  • Optional: yes

  • Common uses: required User-Agent values, organization or account identifiers, or other headers some APIs expect

model_providers:
  - model: moonshotai/kimi-for-coding
    access_key: $MOONSHOTAI_API_KEY
    base_url: https://api.kimi.com/coding/v1
    headers:
      User-Agent: "KimiCLI/1.3"

The example below includes this and other provider options in context.

Plano Configuration - Full Reference
  1# Plano Gateway configuration version
  2version: v0.4.0
  3
  4# External HTTP agents - API type is controlled by request path (/v1/responses, /v1/messages, /v1/chat/completions)
  5agents:
  6  - id: weather_agent # Example agent for weather
  7    url: http://localhost:10510
  8
  9  - id: flight_agent # Example agent for flights
 10    url: http://localhost:10520
 11
 12# MCP filters applied to requests/responses (e.g., input validation, query rewriting)
 13filters:
 14  - id: input_guards # Example filter for input validation
 15    url: http://localhost:10500
 16    # type: mcp (default)
 17    # transport: streamable-http (default)
 18    # tool: input_guards (default - same as filter id)
 19
 20# LLM provider configurations with API keys and model routing
 21model_providers:
 22  - model: openai/gpt-4o
 23    access_key: $OPENAI_API_KEY
 24    default: true
 25
 26  - model: openai/gpt-4o-mini
 27    access_key: $OPENAI_API_KEY
 28
 29  - model: anthropic/claude-sonnet-4-0
 30    access_key: $ANTHROPIC_API_KEY
 31
 32  - model: mistral/ministral-3b-latest
 33    access_key: $MISTRAL_API_KEY
 34
 35  - model: groq/llama-3.3-70b-versatile
 36    access_key: $GROQ_API_KEY
 37
 38  # passthrough_auth: forwards the client's Authorization header upstream instead of
 39  # using the configured access_key. Useful for LiteLLM or similar proxy setups.
 40  - model: openai/gpt-4o-litellm
 41    base_url: https://litellm.example.com
 42    passthrough_auth: true
 43
 44  # Custom/self-hosted endpoint with explicit http_host override
 45  - model: openai/llama-3.3-70b
 46    base_url: https://api.custom-provider.com
 47    http_host: api.custom-provider.com
 48    access_key: $CUSTOM_API_KEY
 49
 50  # headers: optional map of extra HTTP headers sent on upstream requests (after auth).
 51  # Use for provider-specific requirements such as User-Agent, org IDs, or account headers.
 52  - model: moonshotai/kimi-for-coding
 53    access_key: $MOONSHOTAI_API_KEY
 54    base_url: https://api.kimi.com/coding/v1
 55    headers:
 56      User-Agent: "KimiCLI/1.3"
 57
 58# Model aliases - use friendly names instead of full provider model names
 59model_aliases:
 60  fast-llm:
 61    target: gpt-4o-mini
 62
 63  smart-llm:
 64    target: gpt-4o
 65
 66# routing_preferences: top-level list that tags named task categories with an
 67# ordered pool of candidate models. Plano's LLM router matches incoming requests
 68# against these descriptions and returns an ordered list of models; the client
 69# uses models[0] as primary and retries with models[1], models[2]... on 429/5xx.
 70# Requires overrides.llm_routing_model to point at Plano-Orchestrator (or equivalent).
 71# Each model in `models` must be declared in model_providers above.
 72# selection_policy is optional: {prefer: cheapest|fastest|none} lets the router
 73# reorder candidates using live cost/latency data from model_metrics_sources.
 74routing_preferences:
 75  - name: code generation
 76    description: generating new code snippets, functions, or boilerplate based on user prompts or requirements
 77    models:
 78      - anthropic/claude-sonnet-4-0
 79      - openai/gpt-4o
 80      - groq/llama-3.3-70b-versatile
 81  - name: code review
 82    description: reviewing, analyzing, and suggesting improvements to existing code
 83    models:
 84      - anthropic/claude-sonnet-4-0
 85      - groq/llama-3.3-70b-versatile
 86    selection_policy:
 87      prefer: cheapest
 88
 89# model_metrics_sources: external catalogs the router reads to reorder candidate
 90# models for selection_policy.prefer. A `cost` source ranks `prefer: cheapest`;
 91# a `latency` source ranks `prefer: fastest`. Both are optional.
 92model_metrics_sources:
 93  # Cost catalog. provider: models.dev | digitalocean (default url per provider).
 94  - type: cost
 95    provider: models.dev
 96    url: https://models.dev/api.json   # optional; omit to use the provider default
 97    refresh_interval: 3600             # optional, seconds
 98    model_aliases:                     # optional: catalog key -> Plano model name
 99      openai/gpt-oss-120b: openai/gpt-4o
100  # Latency catalog (Prometheus). Used for selection_policy.prefer: fastest.
101  - type: latency
102    provider: prometheus
103    url: http://prometheus:9090
104    query: avg by (model_name) (rate(plano_llm_latency_seconds_sum[5m]))
105    refresh_interval: 60
106
107# HTTP listeners - entry points for agent routing, prompt targets, and direct LLM access
108listeners:
109  # Agent listener for routing requests to multiple agents
110  - type: agent
111    name: travel_booking_service
112    port: 8001
113    router: plano_orchestrator_v1
114    address: 0.0.0.0
115    agents:
116      - id: rag_agent
117        description: virtual assistant for retrieval augmented generation tasks
118        input_filters:
119          - input_guards
120
121  # Model listener for direct LLM access
122  - type: model
123    name: model_1
124    address: 0.0.0.0
125    port: 12000
126    timeout: 30s          # Request timeout (e.g. "30s", "60s")
127    max_retries: 3        # Number of retries on upstream failure
128    input_filters:        # Filters applied before forwarding to LLM
129      - input_guards
130    output_filters:       # Filters applied to LLM responses before returning to client
131      - input_guards
132
133  # Prompt listener for function calling (for prompt_targets)
134  - type: prompt
135    name: prompt_function_listener
136    address: 0.0.0.0
137    port: 10000
138
139# Reusable service endpoints
140endpoints:
141  app_server:
142    endpoint: 127.0.0.1:80
143    connect_timeout: 0.005s
144    protocol: http        # http or https
145
146  mistral_local:
147    endpoint: 127.0.0.1:8001
148
149  secure_service:
150    endpoint: api.example.com:443
151    protocol: https
152    http_host: api.example.com  # Override the Host header sent upstream
153
154# Optional top-level system prompt applied to all prompt_targets
155system_prompt: |
156  You are a helpful assistant. Always respond concisely and accurately.
157
158# Prompt targets for function calling and API orchestration
159prompt_targets:
160  - name: get_current_weather
161    description: Get current weather at a location.
162    parameters:
163      - name: location
164        description: The location to get the weather for
165        required: true
166        type: string
167        format: City, State
168      - name: days
169        description: the number of days for the request
170        required: true
171        type: int
172    endpoint:
173      name: app_server
174      path: /weather
175      http_method: POST
176    # Per-target system prompt (overrides top-level system_prompt for this target)
177    system_prompt: You are a weather expert. Provide accurate and concise weather information.
178    # auto_llm_dispatch_on_response: when true, the LLM is called again with the
179    # function response to produce a final natural-language answer for the user
180    auto_llm_dispatch_on_response: true
181
182# Rate limits - control token usage per model and request selector
183ratelimits:
184  - model: openai/gpt-4o
185    selector:
186      key: x-user-id       # HTTP header key used to identify the rate-limit subject
187      value: "*"           # Wildcard matches any value; use a specific string to target one
188    limit:
189      tokens: 100000       # Maximum tokens allowed in the given time unit
190      unit: hour           # Time unit: "minute", "hour", or "day"
191
192  - model: openai/gpt-4o-mini
193    selector:
194      key: x-org-id
195      value: acme-corp
196    limit:
197      tokens: 500000
198      unit: day
199
200# Global behavior overrides
201overrides:
202  # Threshold for routing a request to a prompt_target (0.0–1.0). Lower = more permissive.
203  prompt_target_intent_matching_threshold: 0.7
204  # Trim conversation history to fit within the model's context window
205  optimize_context_window: true
206  # Use Plano's agent orchestrator for multi-agent request routing
207  use_agent_orchestrator: false
208  # Connect timeout for upstream provider clusters (e.g., "5s", "10s"). Default: "5s"
209  upstream_connect_timeout: 10s
210  # Path to the trusted CA bundle for upstream TLS verification
211  upstream_tls_ca_path: /etc/ssl/certs/ca-certificates.crt
212  # Model used for intent-based LLM routing (must be listed in model_providers)
213  llm_routing_model: Plano-Orchestrator
214  # Model used for agent orchestration (must be listed in model_providers)
215  agent_orchestration_model: Plano-Orchestrator
216  # Disable agentic signal analysis (frustration, repetition, escalation, etc.)
217  # on LLM responses to save CPU. Default: false.
218  disable_signals: false
219
220# Model affinity — pin routing decisions for agentic loops
221routing:
222  session_ttl_seconds: 600    # How long a pinned session lasts (default: 600s / 10 min)
223  session_max_entries: 10000  # Max cached sessions before eviction (upper limit: 10000)
224  # session_cache controls the backend used to store affinity state.
225  # "memory" (default) is in-process and works for single-instance deployments.
226  # "redis" shares state across replicas — required for multi-replica / Kubernetes setups.
227  session_cache:
228    type: memory              # "memory" (default) or "redis"
229    # url is required when type is "redis". Supports redis:// and rediss:// (TLS).
230    # url: redis://localhost:6379
231    # tenant_header: x-org-id  # optional; when set, keys are scoped as plano:affinity:{tenant_id}:{session_id}
232
233# State storage for multi-turn conversation history
234state_storage:
235  type: memory            # "memory" (in-process) or "postgres" (persistent)
236  # connection_string is required when type is postgres.
237  # Supports environment variable substitution: $VAR or ${VAR}
238  # connection_string: postgresql://user:$DB_PASS@localhost:5432/plano
239
240# Input guardrails applied globally to all incoming requests
241prompt_guards:
242  input_guards:
243    jailbreak:
244      on_exception:
245        message: "I'm sorry, I can't help with that request."
246
247# OpenTelemetry tracing configuration
248tracing:
249  # Random sampling percentage (1-100)
250  random_sampling: 100
251  # Include internal Plano spans in traces
252  trace_arch_internal: false
253  # gRPC endpoint for OpenTelemetry collector (e.g., Jaeger, Tempo)
254  opentracing_grpc_endpoint: http://localhost:4317
255  span_attributes:
256    # Propagate request headers whose names start with these prefixes as span attributes
257    header_prefixes:
258      - x-user-
259      - x-org-
260    # Static key/value pairs added to every span
261    static:
262      environment: production
263      service.team: platform