amux/browser-methods.toml at main · mixpeek/amux · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
# amux browser methods: where a browser route runs, and which agents can drive it.
#
# Browser use keeps improving, and new methods should arrive as configuration
# plus a small adapter, never as a rewrite of the route. Two kinds live here:
#
# BACKENDS are where the browser runs. A route tries them in `ladder` order and
# moves to the next only when one cannot serve the request. Today:
#   amux  the server's own Chrome for a saved profile (joins the running
#         browser in a tab of its own when another session holds the profile)
#   cdp   a headless copy of the selected Chrome profile (Browser route settings)
#   cua   the sandboxed desktop (only when CUA is enabled for the profile)
#
# AGENTS are methods that take a natural-language goal and drive whichever
# backend the route is on, through the same state, screenshot and action verbs
# a worker uses: `amux browser route goal '{"task":"...","agent":"<id>"}'`.
# Each agent names a `protocol`, implemented by an adapter module at
# scripts/browser-agents/<protocol>.mjs (or `adapter = "/abs/path.mjs"` for one
# of your own). An adapter is two pure functions: `request(step)` builds the
# model call from the screenshot, task and history, and `parse(reply, step)`
# turns the model's answer into one action or a final answer. The loop, the
# browser, the trajectory log and the safety rules are shared.
#
# To add a method: write its adapter, add an [[agent]] here, point `endpoint`
# at a server that speaks its API, and set enabled = true. To change this on
# one machine only, put the same keys in ~/.amux/browser-methods.toml: an
# [[agent]] with an existing id replaces its fields, a new id adds one, and a
# top-level `ladder` replaces the order.
#
# Screenshots go to the agent's endpoint. A loopback endpoint keeps them on
# this machine; a remote one sends what the browser shows (signed-in pages
# included) to that host, which is why every agent starts disabled.

version = 1
ladder = ["amux", "cdp", "cua"]

[[backend]]
id = "amux"
description = "The amux-managed Chrome for the saved profile; a second session joins it in its own tab."

[[backend]]
id = "cdp"
description = "Headless copy of the Chrome profile chosen in Browser route settings."

[[backend]]
id = "cua"
description = "The sandboxed desktop (amux computer), when CUA is enabled for the profile."

# OpAgent (codefuse-ai, Apache-2.0): a vision-language web agent that reads a
# screenshot and answers with one tool call per step, coordinates on a
# 0-1000 grid. Single-model mode, as published in opagent_single_model/.
# Serve the INT4 model with llama.cpp's OpenAI-compatible server:
#   llama-server --model OpAgent.Q4_K_M.gguf --mmproj mmproj-f16.gguf --port 18080 --ctx-size 4096
# It needs about 20 GB of memory: check `amux resources` first. Setup notes:
# docs/design/browser-methods.md.
[[agent]]
id = "opagent"
protocol = "opagent-v1"
title = "OpAgent single-model (codefuse-ai)"
endpoint = "http://127.0.0.1:18080/v1/chat/completions"
model = "OpAgent-32B-Q4"
enabled = false
max_steps = 15
timeout_s = 180
temperature = 0.0
max_tokens = 1024
history_steps = 2
source = "https://github.com/codefuse-ai/OpAgent"