Local-model defaults through CLIProxyAPI, shared coding skills, and MCP tools for code navigation and web research.
| Path | Purpose |
|---|---|
opencode.jsonc |
Models, providers, MCP servers, and permissions |
dcp.jsonc |
Dynamic Context Pruning settings |
tui.json |
Theme and TUI settings |
AGENTS.md |
Global behavior and tool-selection rules |
agent/ |
Orchestrator, reviewer, docs, and test agents |
skill/ |
Client-specific skills and links to shared skills |
plugin/ |
Active local plugins |
scripts/ |
Configuration and integration checks |
retired/ |
Inactive plugin source |
Restart OpenCode after changing configuration, agents, skills, or plugins.
The configuration and V1 runtime use these process environment variables:
| Variable | Required | Purpose |
|---|---|---|
CLIPROXYAPI_API_KEY |
Yes | Proxy authentication |
FIRECRAWL_API_URL |
Yes for Firecrawl | Self-hosted API endpoint |
FIRECRAWL_API_KEY |
No | Authentication if the Firecrawl instance requires it |
OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX |
Yes for full output budgets | Set to 131072 to lift V1's default 32,000-token ceiling |
Export them in the environment that launches OpenCode. For this lab:
export FIRECRAWL_API_URL=http://172.16.8.179:3002
export OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=131072
# CLIPROXYAPI_API_KEY must already be exported by your credential setup.
opencodeA variable stored only in Pi's agent/.env is not exported to OpenCode. Desktop launches must inherit the variables too.
The llamacpp provider uses CLIProxyAPI at https://cliproxyapi.scherreik.com/v1 through the Responses API. The provider name is retained for existing agent references.
The main model is qwen3.6-27b; small/explore uses qwen3.6-35b-a3b. Both proxy aliases route to MTP builds. The catalog contains the proxy's advertised local aliases, with reasoning, modality, and context metadata from its configuration. Output budgets use publisher recommendations and selected coding examples where available, with a 32,768-token fallback; they are client generation budgets, not model maxima. Cloud models are not included in this provider's catalog.
Context limits use the matching upstream model configs' defaults. The model limits audit records all 26 aliases, primary sources, and optional extended windows. The output limits audit documents generation defaults and recommended budgets separately from hard model limits.
V1 also caps requests at 32,000 tokens unless OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX is exported. The value above permits every selected budget while each model's lower limit.output still applies. See the version-specific source and runtime flag.
Custom agents live in agent/*.md. Select the orchestrator as a primary agent; reviewer, docs, and test are subagents. Built-in agents remain available. The config overrides the explore model and permits the plan agent to edit Markdown files.
OpenCode discovers shared skills from ~/.agents/skills as well as skill/. Skills need name and description frontmatter. Use opencode debug skill to inspect discovery.
The local code-navigation skill describes Grepika and Tilth usage. The local unslop skill provides writing guidance. AGENTS.md holds the shorter tool-selection rules and the no-auto-commit policy.
The local technical-writing skill packages Pstack's Diátaxis, Google developer style, STE, and Global English standard at the same pinned revision as Pi. Its body is unchanged. License and provenance live beside SKILL.md. Global writing instructions and the docs agent require loading technical-writing and unslop through OpenCode's native skill tool. Pi's skill-preload frontmatter and /skill: invocation syntax are not used here. This controls writing style without widening the docs agent's library-documentation scope or enabling shell access.
| Server | Purpose |
|---|---|
| Grepika | Code search, outlines, references, and directory trees |
| Tilth | Symbol definitions, callers, and dependency analysis |
| Context7 | Library documentation |
| gh_grep | Public code search through grep.app |
| Firecrawl | Self-hosted web search, scraping, URL mapping, and crawling |
Chrome DevTools and Next DevTools entries are disabled.
Requires Node.js 22 or newer. OpenCode launches npx -y firecrawl-mcp over stdio, using the endpoint and optional key from the environment.
Keep FIRECRAWL_API_URL exported and pointed at your self-hosted service. Without it, upstream can default to Firecrawl Cloud.
The lab instance currently needs no API key. Leave FIRECRAWL_API_KEY unset unless authentication is enabled. OAuth-token inheritance remains disabled.
Enabled tools:
firecrawl_firecrawl_searchfirecrawl_firecrawl_scrapefirecrawl_firecrawl_mapfirecrawl_firecrawl_crawlfirecrawl_firecrawl_check_crawl_status
OpenCode names MCP tools <server>_<tool>, hence the repeated prefix. Other Firecrawl tools are disabled.
Crawls require approval and finite page/depth limits. Keep external links and subdomains disabled unless requested. The crawl tool waits for completion; the status tool reads an existing job. The request timeout is 120 seconds. A timeout or client interruption does not automatically cancel the remote crawl, so avoid unattended or long-running crawls.
- OpenSlimEdit shortens tool descriptions and read output, and accepts line ranges such as
55-64inedit.oldString. Adapted from Mark Erikson's OpenCode config. - DCP provides dynamic context pruning, configured in
dcp.jsonc.
The old Cachebro bridge is inactive under retired/. OpenCode 1.18.3 does not require it. Read current file contents before editing, using built-in read.
Configuration, MCP discovery, and plugin tests, without model inference:
opencode mcp list
node --experimental-strip-types --test ~/.config/opencode/scripts/verify-plugins.test.mjs
python3 ~/.config/opencode/scripts/verify-firecrawl.pyLive search, scrape, map, and one-page crawl with status readback:
python3 ~/.config/opencode/scripts/verify-firecrawl.py --liveThe live check makes web requests without an approval prompt. It does not test model tool selection or the conversational approval UI.
OpenCode uses native question, task, and todo tools. Pi's managed background-terminal tools, JavaScript workflows, A2A integration, and full custom-agent roster are not installed here. Firecrawl's MCP integration also lacks Pi's automatic crawl cancellation.