Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/mvagnon/opencode-mcp/agents-mdgit clone --depth 1 https://github.com/mvagnon/opencode-mcpWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01107 | $0.01107 |
| Opus 5 | $0.00553 | $0.00553 |
| Sonnet 5 | $0.00221 | $0.00221 |
| Haiku 4.5 | $0.00111 | $0.00111 |
Grade A, and why
opencode-mcp AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 65 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md — opencode-mcp
MCP stdio server (TypeScript, ESM) exposing one tool, ask_codebase: read-only Q&A over GitHub repositories, answered by a headless opencode run inside a locally managed checkout.
Commands
npm run lint # eslint (flat config, eslint.config.mjs)
npm run typecheck # tsc --noEmit — run after every change
npm run build # emit dist/ (the shipped artifact)
npm test # node:test on src/*.test.ts (runs .ts natively, no build)
CI (.github/workflows/ci.yml) runs exactly these four on every push/PR — all must pass.
Architecture
One module per responsibility, all under src/, kebab-case filenames:
| Module | Responsibility |
|---|---|
index.ts |
Bootstrap: create McpServer, register the tool, connect stdio transport |
ask-codebase.ts |
Tool registration, input validation, opencode run + retry, output shaping |
get-repo.ts |
Deterministic clone/fetch/hard-resync (ensureRepo), fetch policy (needsFetch) |
manifest.ts |
JSON manifest of checkouts, atomic writes, mutex-guarded |
proc.ts |
Subprocess runner: hard timeout, 15 s heartbeat, abort-signal kill |
config.ts |
All env-driven configuration, resolved once at startup |
mutex.ts |
Minimal FIFO async mutex |
logger.ts |
Timestamped stderr logging |
Call flow: ask_codebase → ensureRepo (fetch lock → manifest → getRepo if the policy requires) → run(opencode run …) with cwd = checkout → truncated answer + notes.
Hard invariants — do not break
- stdout is the MCP protocol channel. Never
console.log; log vialogger.ts(stderr) only. - The tool is READ-ONLY. The prompt preamble in
buildPromptforbids writes; never add write/exec capabilities to the tool surface. - The manifest format is a shared contract. Keys
dir,branch,fetched_at(ISO 8601) underowner/repo— previous implementations wrote the same file; keep it compatible, keep writes atomic (tmp + rename). - Fetching is deterministic. All user-facing fetch failures go through
GetRepoErrorand are returned verbatim; never let raw exceptions leak to the client as generic errors. - Long subprocesses must heartbeat. Any run that can exceed the client's timeout takes the
heartbeatcallback (progress every 15 s, no-op without a clientprogressToken). - Cancellation kills children. Thread
extra.signalinto everyrun()call; an aborted request must never leave a git or opencode process behind. - Error results use
isError: truewith the message as text content, so the calling LLM sees them.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 65 lines · 1,107 tokens per session scan A 61a3ab1cc659
opencode-mcp AGENTS.md is an instructions file published in the GitHub repository mvagnon/opencode-mcp (0 stars, last pushed 1mo ago), licensed MIT. It adds 1,107 tokens to every session, about $0.0055 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
ebay-mcp AGENTS.md
Instructions for YosefHayim/ebay-mcp, covering agents.md, what this is, instruction sources, validation commands and module map (src/).
taiwan-ecommerce-toolkit CLAUDE.md
Instructions for Moksa1123/taiwan-ecommerce-toolkit, covering claude.md, project overview, available scripts, test api connectivity and validate invoice data.
sparrowhawk AGENTS.md
Instructions for bacpop/sparrowhawk, covering agents.md — sparrowhawk, what this is, getting a working checkout, two facts that will otherwise cost you an hour and layout.
ship-it AGENTS.md
Instructions for LunkiBR/ship-it, covering agents.md, project context, repository layout, terminology and key components.
ScalaSemantic AGENTS.md
Instructions for MercurieVV/ScalaSemantic, covering agents.md instructions and quick start.
agentdir AGENTS.md
Instructions for jstxn/agentdir: Use AgentDir as the local flight recorder for coding-agent work in this repository.