Providers

PanCode discovers and manages LLM providers automatically. It supports local inference engines and cloud API providers, routing models by capability and cost.

Local Engines

PanCode discovers three types of local inference engines at boot:

Engine	Default Port	Protocol
LM Studio	1234	OpenAI-compatible API
Ollama	11434	Ollama native API
llama.cpp (llama-server)	8080	llama.cpp API

Discovery Process

On startup, PanCode probes each known service on each configured machine:

Sends a lightweight health probe (500ms timeout for first tier, 1000ms for second)
Lists available models from responding endpoints
Matches models against the knowledge base for capability metadata
Caches results in ~/.pancode/model-cache.yaml

Boot Performance

Boot Mode	Duration	Behavior
Cold (first run or `--rediscover`)	~1150ms	Probes all endpoints, writes cache
Warm (cached)	~120ms	Reads cache, zero network I/O

On warm boot, PanCode loads from cache and starts a background refresh after the shell is interactive. The refreshed data takes effect on the next boot.

Configuring Machines

By default, PanCode only scans localhost. To discover engines on other machines, set PANCODE_LOCAL_MACHINES:

# In .env or your shell
PANCODE_LOCAL_MACHINES=mini=192.168.86.141,dynamo=192.168.86.143

Format: name1=address1,name2=address2

PanCode probes every combination of machine and service. With 3 machines and 3 services, it probes 9 endpoints.

Provider IDs

Each discovered endpoint gets a provider ID in the format machine-engine:

localhost-ollama
localhost-lmstudio
mini-ollama
dynamo-lmstudio
dynamo-llamacpp

Model References

Models are referenced as provider/model-id:

localhost-ollama/llama3.2
dynamo-lmstudio/qwen2.5-coder-32b
mini-llamacpp/codestral

Use this format everywhere: --model flag, PANCODE_MODEL env var, agent specs, and the /models command.

Cloud API Providers

Set the appropriate API key in your environment:

export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

API provider registration is minimal in the current release. Cloud models are available through the model registry once API keys are configured.

Model Resolution

PanCode resolves the orchestrator model using this priority:

--model CLI flag
PANCODE_MODEL environment variable
.pancode/default-model file (first line)
preferredModel from settings
First available model from discovery

If no model resolves, PanCode starts in degraded mode and surfaces the issue in the shell.

Model Cache

Discovery results are cached at ~/.pancode/model-cache.yaml. This file contains merged model profiles with capability metadata from the knowledge base.

Force Rediscovery

pancode --rediscover    # Cold boot, ignore cache

View Models

/models                 # List all visible models
/models provider/model  # Switch to a specific model

Provider Health

PanCode tracks provider health through the resilience system:

Backoff: Repeated failures trigger exponential backoff
Rate limiting: 429 responses trigger dedicated backoff
Recovery: Successful dispatches reset backoff state

Use /doctor to check provider health status.

Model Knowledge Base

PanCode includes an offline knowledge base with metadata for known models:

Capability classification (coding, reasoning, general)
Parameter counts and context window sizes
Performance tiers (frontier, mid, small)
Cost estimates per provider

The knowledge base lives in the models/ directory of the PanCode installation. It matches discovered models against known entries to provide accurate routing.

Sampling Presets

Agents reference sampling presets that control generation parameters:

Preset	Use Case
`general`	Conversation and analysis
`coding`	Code generation and editing

Agent class profiles (orchestrator, worker, scout) further refine temperature, top-p, and top-k values independently of sampling presets.