Evaluate Anvil in Ten Minutes

Documentation and fixture expectations checked against Anvil 0.23.0 and managed Bifrost 0.8.6 on 2026-07-21. Complete the journey below to verify your editor, provider, and platform.

This journey uses Zed and a checked-in Rust fixture. It proves that an ACP client can start Anvil, select a model, stream a managed-Bifrost tool call, enforce a permission mode, apply one approved edit, and report session context.

It does not measure model quality, large-repository performance, every provider, session recovery in every client, or complete sandbox security.

1. Get Anvil and the Fixture

Follow Install Anvil and confirm:

/absolute/path/to/anvil --version

Clone the repository and open the fixture:

git clone https://github.com/BrokkAi/anvil.git
cd anvil/docs/fixtures/ten-minute-evaluation
zed . # if the Zed CLI is installed

Alternatively, open that fixture directory from Zed’s graphical interface.

The fixture is intentionally small:

pub fn greeting(name: &str) -> String {
    format!("Hello, {name}!")
}

pub fn welcome(name: &str) -> String {
    greeting(name)
}

The greeting definition starts at src/lib.rs:1; its call from welcome is at src/lib.rs:6.

2. Connect Zed

Add an agent server entry using the absolute released-binary path from the Zed guide. Select the newly added Anvil agent, then start a fresh Agent Panel session after changing the configuration.

If Anvil finds a working model, it reports that it is ready. Otherwise it starts setup. Existing Codex credentials are discovered automatically; run this if sign-in is needed:

/setup codex

Use /setup choose if more than one provider is ready. Other providers are covered in Model Providers and Setup.

The first Bifrost-backed prompt may download the Anvil release’s pinned, checksum-verified Bifrost binary. That cold start requires network access and adds installation time; later sessions use the local cache.

3. Prove Read-Only Code Intelligence

Set the session Permission selector to readOnly, then send this exact prompt:

Use Bifrost search_symbols to find greeting, get_symbol_sources to read its definition, and scan_usages_by_reference to find its caller. Report the exact paths and lines. Do not edit files and do not substitute text search.

Success requires all of these signals:

• the Agent Panel shows Bifrost tool cards for the requested operations;
• the answer identifies greeting at src/lib.rs:1;
• the answer identifies the call from welcome at src/lib.rs:6; and
• git diff -- src/lib.rs is empty.

A plausible answer without visible Bifrost calls does not prove the managed MCP path. A file read or grep result does not substitute for the requested symbol and usage tools.

4. Prove the Permission Boundary

Change the Permission selector to default, then send:

Change only the greeting text from `Hello, {name}!` to `Welcome, {name}!`. Show the resulting diff and make no other changes.

Anvil should request permission for the edit tool. Read/search tools may be auto-approved; approve only the constrained edit request once, then verify:

git diff -- src/lib.rs

The only semantic change should be Hello, {name}! to Welcome, {name}!. Rejecting the request should leave the file unchanged. Restore the fixture when finished:

git restore src/lib.rs

5. Inspect Context

Run:

/context

The report should describe the current session’s context and token estimate. If the client exposes session history, close and resume the session and confirm the prior conversation is present; support for that UI is client-specific.

Troubleshooting

• No agent appears: verify that the command is an absolute path and start a fresh Zed session.
• No model is ready: run /setup codex, start Ollama and use /setup local, or configure another provider.
• Bifrost download fails: check network access and the Anvil config/cache directory described in Data and Trust Boundaries.
• No Bifrost tool cards appear: confirm the prompt names the tools and that the managed server started; ordinary source reading is not the same test.
• The edit happens without a prompt: re-check the live Permission selector. acceptEdits and bypassPermissions intentionally behave differently from default.

Interpret the Result

The evaluation separates runtime facts from model-dependent prose. Tool-card presence, source locations, the permission request, and the resulting diff are the evidence. A polished answer alone is not.

Continue with Permissions and Sandboxing before relying on the evaluation as a security claim, and Tools and Managed Bifrost before treating tool output as complete analysis.