VIA

One CLI for orchestrating across models, across harnesses.

_{Click for the full-quality video. Animation source: docs/assets/via-overview.html.}

VIA runs a prompt on a coding-agent harness and gives you back one structured JSON result, whichever harness ran it.

It is one static Rust binary: a CLI plus a per-user daemon that the CLI starts on demand. You name the harness, model, prompt and sandbox bound explicitly; VIA launches the vendor's own CLI (Claude Code and Codex with your existing login; Pi and OpenCode as set up in the user guide), tracks the session, and returns a result envelope with the final text, state, stop reason, token usage, cost and warnings. Roles and orchestration policy stay with the caller.

Why: every coding agent has its own flags, session model, cancel behaviour and output format. VIA gives one pattern (spawn, resume, cancel, close, wait, result, ...) across all of them. Each route declares what it supports, and VIA refuses an unsupported verb or bound by name instead of faking it.

Supported harnesses

Harness	--harness	Route	Model names
Claude Code	claude	claude-cli	sonnet, opus, haiku or a Claude model ID
Codex	codex	codex-app-server	a Codex model name, e.g. gpt-6-luna
OpenCode	opencode	opencode-serve	provider/id
Pi	pi	pi-rpc	provider/id, e.g. openai/gpt-6-luna

steer is not supported by any route in this release. Linux x86_64 only.

Install

Requires the Rust toolchain pinned in rust-toolchain.toml (installed automatically by rustup).

cargo build --release --locked
# binary: target/release/via

A fully static binary (musl) is built and verified by scripts/build-release.sh; see the user guide. Install and log in to each harness CLI you want to use; VIA never handles credentials.

Quick start

umask 077
# Start a turn in the background, in this directory (or pass --cwd D);
# the receipt carries the session handle.
via spawn --harness claude --model haiku --bound full --network \
  --prompt "List the files in this directory and summarise the README." \
  --background > receipt.json
jq -r .handle receipt.json > handle      # needed for resume/cancel/close
TURN=$(jq -r .turn receipt.json)          # e.g. s_7f3k9q2mzr4c/1
SESSION=$(jq -r .session_id receipt.json)

# Wait up to 10 minutes for the result envelope.
via wait "$TURN" --timeout-ms 600000 | jq '{state, stop_reason, final_text, usage, cost}'

# Ask a follow-up in the same session, then close it.
TURN2=$(via resume "$SESSION" --handle-file handle --prompt "Now count them." | jq -r .turn)
via wait "$TURN2" --timeout-ms 600000 | jq -r .final_text
via close "$SESSION" --handle-file handle

Use via describe --harness <h> --model <m> --bound full --network to see what a route supports before running anything; it starts no agent.

Roadmap

In the pipeline, not yet released:

• An ACP route. One generic route for Zed's Agent Client Protocol, so any harness that speaks ACP, natively (such as gemini --acp or opencode acp) or through a bridge, plugs into VIA without a dedicated adapter. This opens VIA to many more harnesses at once.
• More harnesses: Cursor, Gemini CLI, Devin, Amp, Kimi, Goose, Droid, Cline and others.
• More lifecycle control: steering a running turn, and pause.
• macOS. Help from Mac users is very welcome.
• Leftover-process reporting in the result envelope.

Documentation

• User guide: every verb, models, envelope, exit codes, warnings, errors and known limitations.
• Public API (C1): the JSON-RPC contract the CLI speaks.
• Design record.

License

Apache-2.0. See LICENSE.