Manage providers

Construct routes LLM calls across three tiers: reasoning, standard, and fast. Each tier maps to a single model ID drawn from your configured provider (OpenRouter by default).

List current assignments

construct models

Output:

Current model assignments:
  reasoning  openrouter/anthropic/claude-opus-4
  standard   openrouter/anthropic/claude-sonnet-4-5
  fast       openrouter/google/gemini-flash-2-0

Change a tier

Poll the free catalog to see what's available, then set the model for a tier:

construct models free
construct models set --tier=reasoning --model=openrouter/anthropic/claude-opus-4
construct models set --tier=standard --model=openrouter/anthropic/claude-sonnet-4-5
construct models set --tier=fast --model=openrouter/google/gemini-flash-2-0

construct models set rewrites the per-host model config files (Claude Code settings.json, OpenCode config, etc.) so every harness picks up the new assignment immediately.

Control which models appear in pickers via construct.config.json:

construct config set models.visibility.mode all_configured
construct models list
construct models list --json

For a temporary swap (single session), set the env var instead — edit ~/.config/construct/config.env (preferred env keys — CONSTRUCT_MODEL_* is deprecated but still honored):

CX_MODEL_REASONING=openrouter/anthropic/claude-opus-4
CX_MODEL_STANDARD=openrouter/anthropic/claude-sonnet-4-5
CX_MODEL_FAST=openrouter/google/gemini-flash-2-0

Subscription Bridges (Host-Native Models)

If you have an active subscription to GitHub Copilot or Anthropic (via Claude Code), you can utilize these models in OpenCode without needing additional API keys.

Construct automatically detects active sessions (e.g., via gh auth status) and provisions a local Bridge Service that acts as an OpenAI-compatible proxy.

Use Copilot Models in OpenCode

1. Authenticate with GitHub: gh auth login
2. Start Construct services: construct dev
3. Open OpenCode: opencode
4. The models will be available as github-copilot/gpt-4o, github-copilot/claude-3.5-sonnet, etc.

See the Host-Native Models guide for more details.

Adding External Data Sources

Construct connects to external systems via providers. Only configured repos/projects are focal anchors: everything else is ambient (ignored by auto-fetch, searchable via Rovo).

GitHub Repos

# In ~/.config/construct/config.env
GITHUB_REPOS=hashicorp/project-iverson,hashicorp/cloud-reliability,hashicorp/team-delivery-intelligence

Jira Projects

# In ~/.config/construct/config.env
JIRA_BASE_URL=https://hashicorp.atlassian.net
JIRA_USER_EMAIL=your@email.com
JIRA_API_TOKEN=your-api-token
JIRA_PROJECTS=RLBLT,DI,RRM
JIRA_FETCH_RECENCY_DAYS=30

Verify Provider Connections

construct providers status lists every configured provider with its enabled/health state and circuit-breaker status:

construct providers status

Fetch from a Focal Source

provider_fetch is an MCP tool, not a CLI command. The agent (or any MCP host connected to the Construct MCP server) calls it to pull a focal source into memory:

provider_fetch("project iverson")
provider_fetch("RLBLT")

Broad Search Across All Sources

rovo_search is likewise an MCP tool. It runs Rovo AI search — returns excerpts, does NOT store in memory:

rovo_search("Iverson reliability")

Embedding Models

Construct uses neural embeddings for semantic search. The model is configurable via the dashboard or env vars.

Available Models

ID	Provider	Model	Dimensions	Cost
local	ONNX	Xenova/all-MiniLM-L6-v2	384	Free
openai	OpenAI	text-embedding-3-small	1536	Paid
ollama	Ollama	nomic-embed-text	768	Free (local)
hashing	Local	hashing-bow-v1	256	Free (legacy)

Change Embedding Model

# In ~/.config/construct/config.env
CONSTRUCT_EMBEDDING_MODEL=openai
OPENAI_API_KEY=sk-...

Or via the dashboard Models screen.

List all agents with quality scores

construct optimize --list

Shows each agent's average quality score and trace count: useful for deciding which tier to bump up before a review.

Dry-run prompt optimization

construct optimize cx-engineer --dry-run

Previews prompt changes inferred from low-quality traces without applying them.

Apply prompt optimization

construct optimize cx-engineer

Rewrites the agent's system prompt slice in registry.json based on recurring failure patterns in telemetry traces. Requires CONSTRUCT_TELEMETRY_PUBLIC_KEY and CONSTRUCT_TELEMETRY_SECRET_KEY to be set.

Verify after changes

construct doctor
construct diff

doctor confirms all integrations are healthy. diff shows which agent prompts changed since HEAD.