Adding agent integrations

September 5, 2026 · View on GitHub

Minutes has several agent surfaces. Do not add a new agent by copying an existing integration wholesale. Pick the smallest surface that matches what the host actually supports.

Install the Claude Code plugin

The repository is a Claude Code plugin marketplace. Add it, then install the Minutes plugin:

claude plugin marketplace add silverstein/minutes
claude plugin install minutes@minutes

The plugin starts npx -y minutes-mcp automatically, so its Minutes tools are available without a separate MCP configuration.

MCP server CLI auto-install

When an MCP host starts minutes-mcp without a compatible Minutes CLI, the server installs a checksum-verified release asset. On Apple Silicon (including Node running under Rosetta), it selects minutes-macos-arm64-sherpa.tar.gz, which contains the CLI and its Developer ID-signed sherpa plugin. The server extracts into a temporary directory, confirms the staged CLI runs, and then atomically replaces the installed binary and plugin. Releases from before the archive existed fall back to the checksum-verified bare macOS binary and log that Parakeet is unavailable.

After installation, a default engine = "auto" health check that finds the plugin but no Parakeet model triggers plain minutes setup. That installs the Parakeet model and the Whisper tiny fallback without changing explicit engine choices.

For a host that only reads existing transcripts, set MINUTES_MCP_AUTO_SETUP to "0" in the MCP server's environment to skip speech-model health checks and downloads. The generated minutes-mcp --demo configuration sets this option. The CLI remains required for meeting access rules and can still be installed automatically. Before recording, use the desktop setup or minutes setup to prepare the speech model.

MEETINGS_DIR selects the corpus for MCP recall tools, including consistency reports. The server passes that folder explicitly to minutes consistency --dir <folder> and requires a CLI with the consistency_corpus_root_v1 capability. An older CLI must be updated before it can produce this report; the server does not fall back to a different meeting library.

Surfaces

Surface	Use when	Examples
Raw files	The agent can read `~/meetings/` directly.	Cursor, any local coding agent
MCP server	The host supports MCP tools/resources/prompts.	Claude Desktop, Codex, Gemini CLI, OpenClaw, Hermes-agent
Portable skills	The host discovers Agent Skills-style `.agents/skills` folders.	Codex, Gemini CLI, Pi, OpenClaw, Hermes-agent
Host-specific skills	The host needs a different generated shape.	Claude Code plugin, OpenCode commands
`agent_command` backend	Minutes should call the agent CLI for summaries.	`claude`, `codex`, `opencode`, `pi`, `agent` (Cursor Agent CLI)
OpenAI-compatible model backend	Minutes should call a model API directly, not an agent CLI.	OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway, llama.cpp, vLLM
Routing eval	The agent has a non-interactive prompt mode worth benchmarking.	`npm --prefix tooling/skills run routing:agents -- --agent codex`

Agent hosts vs model providers

Do not treat every AI brand as an agent integration. Minutes has two separate contracts:

Agent hosts run their own agent loop, tool policy, memory, and prompt wrapper. They belong in agent_command only when Minutes can safely invoke them non-interactively and capture stdout.
Model providers expose inference APIs. They belong behind a direct summarization backend, ideally a generic OpenAI-compatible backend with provider presets.

OpenCode is an agent host. It can use many providers internally, but that only covers the agent_command = "opencode" path. It does not replace Minutes supporting direct model backends for users who want summaries without an external coding-agent loop.

Model backend candidates

Prefer one generic OpenAI-compatible backend over one engine per provider. That keeps settings small and lets advanced users bring their own gateway or local runtime.

Backend	Classification	Notes
Ollama	Local model runtime	Already supported directly as `engine = "ollama"`; can also be reached through its OpenAI-compatible endpoint.
llama.cpp / llama-cpp-python	Local model runtime	Support through an OpenAI-compatible `base_url`; do not add as an agent option.
vLLM / LM Studio / LocalAI	Local or self-hosted runtime	Support through the same OpenAI-compatible path when available.
OpenRouter	Cloud model router	Good preset for one-key access to many providers; transcripts leave the machine.
Vercel AI Gateway	Cloud model gateway	Good preset for hosted apps and teams already using Vercel; transcripts leave the machine.
Cloudflare AI Gateway	Cloud model gateway	Good preset for observability, rate limits, caching, retries, and Cloudflare-managed routing; transcripts leave the machine unless the upstream is local/private.

Recommended config shape for future direct backends:

[summarization]
engine = "openai-compatible"
openai_compatible_model = "openai/gpt-4o-mini"
openai_compatible_base_url = "https://gateway.example.com/v1"
openai_compatible_api_key_env = "AI_GATEWAY_API_KEY"

In the desktop app, cloud gateway presets use the same backend with a friendlier secret path: users paste the key in Settings, Minutes stores it in macOS Keychain, and the app hydrates MINUTES_OPENAI_COMPATIBLE_API_KEY for its own summarization calls. Keep provider-specific env vars as the CLI/power-user fallback, not as the default desktop experience.

Checklist

Identify the host contract.
- Can it read files?
- Does it support MCP?
- Does it auto-discover .agents/skills?
- Does it require a host-specific skill tree?
- Does it have a non-interactive CLI mode?
If the host can reuse .agents/skills, do not generate a duplicate tree. Duplicate skill names can create collisions and make the agent less reliable.
If the host needs a generated skill surface, update:
- tooling/skills/schema.ts
- tooling/skills/hosts/
- tooling/skills/compiler/render.ts
- tooling/skills/compiler/compile.ts
- tooling/skills/compiler/check.ts
- tooling/skills/compiler/golden.ts
- generated outputs under the host-specific tree
If the host should be callable from Minutes summarization, update:
- crates/core/src/summarize.rs
- targeted prepare_agent_invocation_* tests
- tauri/src-tauri/src/commands.rs
- tauri/src/index.html
- docs/architecture/config.md
If adding a direct model backend rather than an agent host, update:
- crates/core/src/summarize.rs
- SummarizationConfig in crates/core/src/config.rs
- desktop settings and validation in tauri/src-tauri/src/commands.rs
- tauri/src/index.html
- docs/architecture/config.md
- provider-specific docs only when there are real caveats
If the host should participate in routing evals, update:
- tooling/skills/compiler/agent-routing.ts
- tooling/skills/compiler/agent-routing.test.ts if parsing or unavailable handling changes
Update public and agent-facing docs:
- README.md
- site/app/for-agents/page.tsx
- site/lib/product-surfaces.json
- manifest.json
- docs/architecture/config.md
- docs/<agent>.md when the host has provider-specific caveats
- run node scripts/generate_llms_txt.mjs
Run the relevant gates:
- cargo fmt
- targeted Rust tests for the invocation path
- cargo check -p minutes-app
- npm --prefix tooling/skills run build
- npm --prefix tooling/skills run compile:dry
- npm --prefix tooling/skills run check
- npm --prefix tooling/skills run test
- npm --prefix site run check:llms
- npm --prefix site run build when site pages changed

Current agent classes

Claude Code: host-specific plugin surface plus MCP.
OpenCode: host-specific .opencode/skills and .opencode/commands, plus MCP when configured.
Codex: portable .agents/skills plus MCP.
Gemini CLI: portable .agents/skills plus MCP.
Pi coding agent: portable .agents/skills plus opt-in agent_command = "pi" summarization. No separate .pi/skills tree.
Cursor: raw meeting files and MCP where the host supports it, plus opt-in agent_command = "agent" (Cursor Agent CLI) for non-interactive summarization. See cursor-agent.md.

Experimental runtime watchlist

These are real enough to mention but should not become first-class UI options until a small spike proves the command contract:

Runtime	Why it matters	Integration posture
Goose	Open-source on-machine agent with CLI, API, MCP extensions, and custom OpenAI-compatible provider support.	Spike as `agent_command` only if `goose` has a clean non-interactive run mode for transcript summaries.
Hermes Agent	Persistent personal agent with gateway, memory, skills, browser control, OpenRouter, custom APIs, and local vLLM support.	Treat as an experimental agent host and portable-skills target; verify non-interactive stdout behavior first.
OpenClaw	Local-first personal automation gateway with many messaging channels, tools, and daemon-style routing.	Prefer webhook/notification recipes first; only add `agent_command` after a security review and command-contract spike.
Aider	Open-source terminal pair programmer with broad model support, including OpenRouter.	Coding-focused; verify read-only, non-editing summarization behavior before exposing.
OpenHands	Open-source agent platform and SDK with local and sandboxed deployment modes.	Advanced/platform integration, not a simple dropdown peer to Codex or OpenCode.

When in doubt, prefer the raw file or MCP path first. Add a custom host surface only when the agent cannot consume the existing portable one.