eVi

Delegate to another machine on your LAN

Keep a small, fast model for everyday chat on your laptop, and hand the heavy turns to a bigger eVi running on a GPU box down the hall — no cloud, no shared account, just two eVi instances talking over your LAN.

Uses: config ([tools] federation, [federation] serve, ~/.evi/peers.json) · CLI (evi peer …) · two machines on one trusted network.

How it works

Your laptop POSTs a self-contained task to the peer's /api/federate endpoint; the peer runs it on its own agent and returns the answer text. Transport is plain HTTP with the peer's existing web bearer token — there's no separate trust model and no TLS, so keep peers on a LAN or VPN. Federation is opt-in on both ends: the caller enables a tool, the responder opts in to serving.

On the big box (the one doing the work)

Turn on serving and bind the web server to the LAN so peers can actually reach it:

# ~/.evi/config.toml
[web]
auth_token = "…"      # you're on the LAN now — require a token (evi web-config token rotate)

[federation]
serve = true          # answer POST /api/federate for trusted peers (403 otherwise)
evi web --host 0.0.0.0 --port 8473   # bind the LAN on the federation port

evi web defaults to 127.0.0.1:8000, so both flags matter: --host 0.0.0.0 to leave loopback, and --port 8473 because that's the port peers scan and the http://gpu-box:8473 URL below points at. Then open port 8473 through the firewall. A served task runs non-interactively and is deny-by-default for tools — the peer's agent refuses any tool not already auto-approved, so delegation can't quietly trigger a shell or network tool on the remote box.

On the small box (the one delegating)

Enable the delegate_peer tool and register the peer:

# ~/.evi/config.toml
[tools]
federation = true     # enables the delegate_peer tool (off by default)
evi peer add gpu http://gpu-box:8473 --token <big-box web token>
evi peer list         # shows live reachability + the peer's eVi version/model

Prefer a file? Copy examples/peers.json to ~/.evi/peers.json and fill in each name / url / token (token is optional if the peer needs no auth).

Delegate a task

From the CLI, one shot:

evi peer run gpu "Summarize the architecture of /srv/project and list its top 5 risks."

From a chat, with the tool on, the model calls delegate_peer(peer, task) itself (and list_peers first, to pick the peer whose model fits the job — e.g. route a vision task to the peer whose model has vision).

Discover peers on the network

evi peer scan            # sweep the local /24 for running eVi instances (~1-2 s)
evi peer scan --port 8000

If a scan finds nothing with the firewall open, the peer is almost certainly bound to 127.0.0.1 only — relaunch it with evi web --host 0.0.0.0 --port 8473. evi peer scan warns you when this machine has that problem.

Distribute a whole task list

Once peers are registered, evi team run --peers (alias --distribute) fans a shared task list across reachable peers plus local — real cross-machine parallelism instead of teammates serialising on one GPU. A peer that drops mid-run falls back to running its task locally, so the team never stalls.

The full field reference (serving, A2A, the Peers panel, security notes) is in docs/features/agents.md.

Where to go next

Edit on GitHub ↗