Amp 2026-08-04: File Attachments and Event-Driven Cloud Orbs#
Part of the Amp Version Tracker series. | ampcode.com | News
Build 0.0.1785875347 gives plugins and agents a real attachment upload API and fills in the plumbing behind Amp's hosted "orb" runtime, three weeks of changelog work landing as binary surface since the last tracked build 0.0.1784046309.
New Environment Variables#
The clearest additions cluster around running an agent inside a remote sandbox (an "orb") instead of on your laptop. Each of these is wired into an executor code path in the binary, not just declared.
| Variable | What It Does |
|---|---|
AMP_GIT_PASSWORD | Feeds a git credential helper so a hosted orb can clone the thread's repository. The binary builds a helper inline (echo "password=$AMP_GIT_PASSWORD") and clones from /git/<path> on the Amp URL as user x-amp-user. This is how an orb gets the code without your local SSH keys. |
AMP_GITHUB_BOOTSTRAP_TOKEN | Sits on the orb's environment passthrough allowlist alongside AMP_API_KEY. A GitHub token forwarded into the sandbox so the remote agent can reach GitHub on your behalf. |
AMP_DIRECT_TERMINAL | Set to 1 to attach your local terminal directly to the sandbox executor. The gate fires only when the executor type is sandbox and the sandbox's AMP_THREAD_ID matches the thread you are viewing, so the connection is scoped to one running orb. |
AMP_ORB_WORKLOAD_SLICE | On Linux, pins the orb's child processes into a systemd cgroup slice. The value must be amp-workload.slice or the launch throws, so an orb's compute stays inside one resource-bounded unit. |
Two more identity values, AMP_WORKSPACE_ID and AMP_USER_ID, ride the same passthrough allowlist into the sandbox, matching the shared-orb "Multiplayer" work from the July 22 changelog[1] that lets a team share control of one orb. Alongside them, a set of AMP_PORTAL_HAIRPIN_* certificate variables and a generated "Amp Orb Portal Hairpin CA" (installed with an /etc/hosts edit) supply the TLS and DNS routing an orb needs to receive inbound requests, the networking under "Event Driven Orbs" from July 23[1].
Attachment Uploads and the Model Roster#
The August 4 "Attach Anything"[1] post ships as a concrete plugin API: amp.attachments.upload({ data, mimeType }) returns a hosted { url }. The in-binary example uploads a screenshot as a Uint8Array and hands back an image block, so an agent can now push videos, logs, PDFs, or datasets into a thread and reference them by URL instead of pasting text.
On the model side, claude-opus-5 enters the routing table, and three cheaper Google tiers appear (gemini-3.1-flash-lite, gemini-3.5-flash-lite, gemini-3.6-flash). That tracks the July 29 "Who Cares About the Model?"[1] note, where the team swapped the default model overnight and nobody noticed, Amp keeps the choice on its side of the wire.
Supporting the orb runtime, a /usr/local/bin/amp-run-workload wrapper launches scoped workloads, and amp-snapshot-index / amp-snapshot-objects build temporary git-object snapshots to sync your working tree into the orb without a push. The build also opens local sockets for an amp-desktop-v1 and amp-terminal-v1 endpoint (~/.amp/desktop, ~/.amp/terminal), IPC surface for connecting the CLI to a desktop process.
What This Tells Us#
The whole diff points one direction: Sourcegraph is moving the agent off your machine. Claude Code, Codex CLI, Gemini CLI, and OpenCode all run the model loop locally against your files. Amp's orb is a hosted sandbox that clones your repo with its own credentials (AMP_GIT_PASSWORD, AMP_GITHUB_BOOTSTRAP_TOKEN), runs inside a systemd resource slice, carries a workspace and user identity for shared control, and exposes a network endpoint so outside events can wake it. That is closer to a cloud CI runner for agents than to a terminal tool.
Attachments and the model changes reinforce the same posture. Uploading arbitrary files to a hosted URL only makes sense when the agent runs somewhere with storage and network of its own. And quietly adding claude-opus-5 while swapping the default with no announcement is the same "no model picker" philosophy Sourcegraph has held from the start: the harness decides, you get the work. Against competitors still shipping model selectors and local-first loops, Amp is betting that most of the value is in owning the execution environment, not the prompt.
For anyone tracking these builds, the signal is that the orb is graduating from experiment to product surface. The env vars here are load-bearing, not feature flags, credential handling, cgroup scoping, and TLS routing are the parts you build once you expect real workloads to run unattended.
This analysis is conducted for educational and research purposes under fair use principles. All trademarks and software referenced belong to their respective owners. This content is not intended to infringe on any intellectual property rights, circumvent any protections, or encourage unauthorized access to proprietary systems.
Sources#
- Amp News — entries dated July 22, July 23, July 29, and August 4, 2026.