Agent integrations

Add the smallest useful agent surface.

Minutes is deliberately not one integration per agent. Most hosts can use the same local markdown, MCP server, or portable skill pack. Add a custom surface only when the host contract actually requires one.

Agent hosts

Hosts run their own agent loop, prompt wrapper, tools, and memory. Add them to agent_command only after a non-interactive stdout contract is proven.

Model providers

Providers expose inference APIs. OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway, and llama.cpp belong behind a generic OpenAI-compatible backend, not in the agent list.

Choose the surface

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, and prompts.	Claude Desktop, Codex, Gemini CLI
Portable skills	The host discovers Agent Skills-style .agents/skills folders.	Codex, Gemini CLI, Pi
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
OpenAI-compatible backend	Minutes should call a model API directly, not an agent loop.	OpenRouter, Vercel AI Gateway, Cloudflare AI Gateway, llama.cpp
Routing eval	The agent has a non-interactive prompt mode worth benchmarking.	routing:agents -- --agent codex

Model backends

Backend	Class	Posture
Ollama	Local model runtime	Already supported directly; can also be reached through its OpenAI-compatible endpoint.
llama.cpp	Local model runtime	Use an OpenAI-compatible base URL. Do not add it as an agent option.
vLLM / LM Studio / LocalAI	Local or self-hosted runtime	Use the same OpenAI-compatible path when available.
OpenRouter	Cloud model router	Preset candidate for one-key access to many providers.
Vercel AI Gateway	Cloud model gateway	Preset candidate for teams already using Vercel routing, billing, or AI SDK workflows.
Cloudflare AI Gateway	Cloud model gateway	Preset candidate for observability, rate limits, caching, retries, and Cloudflare routing.

Checklist

1.Identify whether the host supports file reads, MCP, .agents skill discovery, host-specific skill discovery, and non-interactive CLI prompts.
2.Reuse .agents/skills when the host already discovers it. Do not add a duplicate generated tree just for symmetry.
3.Add a host-specific skill compiler target only when the host cannot consume the portable skill pack.
4.For summarization backends, update summarize.rs, desktop settings, docs/CONFIG.md, and targeted invocation tests.
5.For model providers, prefer one generic OpenAI-compatible backend with presets over one top-level engine per gateway.
6.For routing evals, update tooling/skills/compiler/agent-routing.ts and run the routing-agent smoke where auth allows.
7.Update README, /for-agents, product-surfaces.json, manifest.json, llms.txt, and any provider-specific docs.
8.Run Rust, skill-tooling, llms, and site gates before shipping.

Current map

Claude Code

Host-specific plugin surface plus MCP.

OpenCode

Host-specific .opencode skills and 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 .pi/skills export.

Cursor and editors

Raw meeting files and MCP where the host supports it.

Experimental runtimes

Goose

Open-source on-machine agent with CLI, API, MCP extensions, and custom OpenAI-compatible provider support.

Hermes Agent

Persistent personal agent with gateway, memory, skills, browser control, OpenRouter, custom APIs, and local vLLM support.

OpenClaw

Local-first personal automation gateway with messaging channels, tools, and daemon-style routing.

Aider

Open-source terminal pair programmer with broad model support, including OpenRouter.

OpenHands

Open-source agent platform and SDK with local and sandboxed deployment modes.

These should stay in docs or recipes until a spike verifies read-only behavior, non-interactive execution, stable stdout, and permission boundaries.

Repo checklist

The full contributor checklist lives in docs/AGENT-INTEGRATIONS.md. It names the exact files to update for compiler hosts, summarization backends, routing evals, and generated public agent docs.