Replies: 1 comment
|
I work on Open Agent API, a separate ACP client that exposes selected agents through SDK-compatible HTTP endpoints. Your split between local ACP and a narrow task API makes sense to me; I would standardize the task API before trying to relay arbitrary ACP messages. Two adapter edges we account for: a resume request needs to be gated on what the agent advertises, and sending session/cancel is not proof the run stopped. Our stdio adapter sends the cancellation first and has a bounded process-stop fallback. On permissions, we map an approval to allow_once, not a persistent grant. I would keep the relay reconnect semantics independent of agent session resume so a reconnect does not accidentally start a second run or replay an approval. For the website-supplied instructions, I would keep site text separate from the owner-controlled policy rather than relying on append order as a trust boundary. Our ACP adapter notes describe the cancellation, approval and resume distinctions. Have you seen any adapters advertise resume but still require a fresh session after the desktop process restarts? |
Uh oh!
There was an error while loading. Please reload this page.
Hi all 👋 I've been building Remote ACP Bridge, an open-source (Apache-2.0) desktop app that connects local ACP agents to websites the user explicitly approves. It currently supports Codex via
codex-acpand Claude Code viaclaude-agent-acp. I'd love feedback from people who know ACP well.The idea: ACP stays local, over stdio, between the desktop app and the agent. A website never speaks raw ACP. It gets a small, allowlisted task API through an outbound-only relay: start a task, stream sanitized
session/updateevents, cancel. The desktop mediates everything:session/request_permissionis always answered locally; the website can only observe itread-only/default), and permissive modes are never selectedfs/*, and that Codex'sread-onlymode can read the whole disk.Websites can also pass
systemPrompt(append mode) and their ownmcpServers(HTTP, restricted to the site's verified domain). That enables things like the flagship example: an issue tracker with "Fix with my agent". The user's own agent fixes the bug locally, runs the tests, and reports back through the tracker's MCP tools.Repo: https://github.com/brstic/remote-acp
Things I'd especially value input on:
systemPrompt: is append mode the right default?Thanks for ACP, it made this possible.
All reactions