Feature request — cost/priority metadata for MCP tools; agents default to the most expensive tool

Version: opencode 1.18.14, Linux

Problem

When an MCP server exposes multiple tiers of the same capability at different prices (e.g., a search API with sonar at $1/$1 per M tokens, sonar-pro at $3/$15, and a deep-research mode at $2/$8 + extra fees), the LLM has no signal to prefer the cheap tier. It defaults to the most capable-sounding — and therefore most expensive — tool.

Concretely:

• Tools are only controllable via permission (allow/ask/deny) and tools (enable/disable). There is no way to express cost, priority, or a "default to this one" preference on a tool.
• Description text is the only discriminator, and models reliably gravitate to "most comprehensive" descriptions even when they say "use ONLY as a last resort."
• In my session, the tools surfaced to the model in the order sonar_deep_research → sonar_reason → sonar_search_pro → sonar_search — the most expensive listed first, despite being defined cheapest-first in the server. Whatever the sort, it's working against cost-awareness.

Suggested solutions (any would help)

1. Cost/priority field on tool definitions — e.g. an optional cost or preference hint (MCP server-provided or config-level) that biases selection and/or sorts cheap-first.
2. Per-agent tool ordering — let config express a preferred order for tools, applied to the tool list the model sees.
3. opencode mcp list surfacing tool count/order — so users can at least see what order the model is presented.
4. A config-level "default tool" per MCP server — e.g. "tools": {"perplexity": {"default": "sonar_search"}}.

FYI — how we're working around it today

We run a single local MCP server (~/.config/opencode/mcp/perplexity/server.js) exposing the four tiers as explicitly-named tools with no auto-escalation — each tool calls exactly the model it's named for, so the tool choice is the cost choice:

• sonar_search → sonar ($1/$1)
• sonar_search_pro → sonar-pro ($3/$15)
• sonar_reason → sonar-reasoning-pro ($2/$8)
• sonar_deep_research → sonar-deep-research ($2/$8+)

(We replaced the third-party perplexity-mcp npm package, which silently auto-escalated based on query keywords — that made cost control impossible.)

The remaining mitigation is a skill (perplexity-cost-aware) that instructs the model to start with sonar_search and escalate only when the cheap answer is insufficient. It works intermittently — it's advisory and only helps if the model loads it before the first call. A structural mechanism in opencode would be far more reliable than description text and skills.

The issue does not name repository files or tests. Start by tracing MCP tool discovery, ordering, configuration handling, and the opencode mcp list entry point; first clarify which proposed behavior is intended. Done should include a defined cost, priority, ordering, or default-tool mechanism with user-visible behavior and coverage for the selected configuration.