Configure

The backend reads one file, cotracer.yaml (mounted at /app/cotracer.yaml), and secrets from environment variables. Unknown keys are rejected at startup, so a typo fails loudly instead of being ignored. Remove a section to turn that integration off — its tools are never built and the model never sees them.

Example

llm:
  provider: anthropic
  model: claude-opus-4-8
  api_key_env: ANTHROPIC_API_KEY

tracing:
  provider: jaeger
  url: http://jaeger-query.internal:16686

logs:
  provider: opensearch
  url: http://opensearch.internal:9200

vcs:
  provider: github
  org: your-org              # cross-repo search scope
  # repo: your-org/default   # fallback when a request names no repo

Secrets never go in this file. Where a provider needs one, the config names the environment variable that holds it.

llm — your model

Key	Meaning
provider	provider id, e.g. anthropic, openai
model	the provider’s model id, e.g. claude-opus-4-8, gpt-…
url	optional — your own endpoint (self-hosted model, internal gateway)
api_key_env	env var holding the key; startup fails if it is unset

api_key_env can name any variable (e.g. LLM_GATEWAY_KEY); the backend passes its value to the provider.

Model choice matters: use the strongest model your policy allows. Small local models guess service names and stop early.

Without an llm section the backend falls back to a small local model — fine for a smoke test, not for real analyses.

tracing

Key	Meaning
provider	provider id, e.g. jaeger
url	your tracing backend’s query endpoint

logs

Key	Meaning
provider	provider id, e.g. opensearch
url	your logs backend’s endpoint

vcs

Key	Default	Meaning
provider		github
url	https://api.github.com	GitHub Enterprise Server: https://<host>/api/v3
org		your GitHub org — where the App is installed; scopes cross-repo search. Set org or repo
repo		default owner/name when a request names none
app_id_env	GITHUB_APP_ID	env var with the App id
app_private_key_env	GITHUB_APP_PRIVATE_KEY	env var with the App private key (PEM contents)
cache_dir	~/.cache/cotracer/repos	local bare clones used for code search
cache_max_gb	20	least-recently-used repos are evicted beyond this
clone_size_cap_mb	500	larger repos are not cloned

Code search and file reads run on local bare clones of your repos, fetched lazily and updated incrementally — keep cache_dir on a persistent volume. Setup of the App itself: Connect GitHub.

Environment variables

Variable	Required	Meaning
COTRACER_API_KEY	yes	clients must send it (Authorization: Bearer …). Unset = auth disabled — never in production
COTRACER_LICENSE_TOKEN	yes*	clt_… license fetch token (*or mount a license file)
your LLM key	per llm	the variable named by llm.api_key_env
GITHUB_APP_ID, GITHUB_APP_PRIVATE_KEY	with vcs	GitHub App credentials
COTRACER_LICENSE_FILE	no	license file path (default /app/cotracer-license.jwt)
COTRACER_LICENSE_URL	no	license server (default https://platform.cotracer.ai)
HTTPS_PROXY, NO_PROXY	no	outbound proxy for the LLM, GitHub and license refresh