Coding agents

Twelve coding agents. Seven are verified end to end — registered, handshaking, listing tools, and answering a live call — on the CLI version current at the last conformance run. Two have not been through one yet: Antigravity is built from the vendor's documented CLI, and Cursor from an integration session against the editor itself. Pi is the one harness Archis does not reach over MCP at all, because Pi has no MCP client: Archis arrives there as the briefing and the skill instead. JetBrains Air, added since, is measured on macOS and not yet on Linux.

A vendor command is always preferred over editing a file: the vendor owns its own config schema, so Archis never has to track it. The MCP harnesses with no scriptable command are edited directly — backed up to <file>.chi-backup first, written through a temporary file and an atomic rename, with every unrelated key preserved.

Bare, chi agents reports every harness it knows: whether its CLI is on this machine, whether Archis's tools are registered with it, whether the hooks are installed, and whether it has been briefed.

With no id, on arranges every harness it found. With an id, only that one. The flags narrow it further when you want one piece — the registration but not the briefing, say — and chi agents off reverses any install Archis performed, leaving every file otherwise as it was.

Every file chi agents on can write for a harness, global and per repository, who writes it and the command that takes it back out. Pick an agent; the URL follows, so this is a page you can send to somebody. The same answer, for the machine you are on rather than in general, is chi config; chi config --catalog prints this table in the terminal.

Claude Code · claude · Anthropic · install: npm i -g @anthropic-ai/claude-code

"Your home" means the files under your own account — ~/ — which apply to every repository you open and are never committed. "This repository" means files at the repository root, which usually are; nothing is written there unless you ask with --project.

Independent of the agent, Archis keeps its own state in two places: a directory per machine, and .chi/ at the root of each repository it has indexed. Nothing in either is sent anywhere; the privacy page has the boundary.

.git/info/exclude rather than .gitignore, on purpose: it does the same job without Archis modifying a file you have committed. And any file Archis edits in place — an instructions file, a harness's JSON or YAML config — is copied first to <file>.chi-backup, once, and the copy is yours.

Generated from chi 0.52.0.

The most expensive mistake with Archis is a connection that works. Registering the tools puts them in front of the model. It does not tell the model they are there, and a model that was never told goes on reading files — the tools sit unused and the session costs exactly what it always did.

There are three channels, in descending order of how much Archis controls them, and no harness has all three:

chi agents on uses every channel the harness supports. chi agents advice reports whether Archis is speaking up before an agent reaches for its own file tools, which is the single check worth running if the numbers are not moving.

Nine tools, and the shape of them is the product: the plural ones exist so that one call answers a question that would otherwise be ten. A turn re-sends the whole conversation, so the number of calls, not the size of any answer, is what a session costs. Start with chi_context: it takes the task in words and returns what the change needs.

That is the default set. chi-mcp --toolset full advertises all 37, including the singular ones, for clients that need them; a wider list invites one call per fact, which is why it is not the default.

Where a registry composes the repository with others in a federated workspace, on a Team or Enterprise plan, every toolset also carries federation_context: chi_context across the repositories of a federated workspace, those that live only on the organization’s CAVS DB Server included: lines from each, with the repository and exact commit they came from, and nothing of a repository the user may not read.

Every one of them has a CLI equivalent, which is the useful thing about the CLI: when an agent gives you an answer you do not believe, you can ask the same question by hand. The reference maps them.

Six things are checked, in the order they can fail:

• The harness CLI is on PATH.
• The harness itself reports chi among its servers — asked of the harness, not read back from the file Archis wrote. Reading our own write proves only that we can write.
• The server binary chi-mcp resolves on PATH.
• The handshake returns a server identity and a protocol version.
• At least one tool is advertised.
• A real tool call returns a real answer.

The last three run the exact command line that was registered, so a registration naming a broken path fails here rather than silently at your next session. The command exits non-zero on failure, so it works in CI.

chi agents off <id> undoes one connection. chi eject --machine undoes all of them, plus the hooks, the briefing and the machine-wide registrations, and touches no repository. Both remove only what is attributably Archis's — and chi config shows you the list before either runs.

chi run "<task>" drives a live model over Archis's tools until it finishes or hits a step limit. It exists for two reasons: to run a task unattended, and to measure Archis against a full-source baseline on the same task — --no-compress is the control arm, and the difference between the two runs is the whole claim.

Writing files and running commands are separate opt-ins — --write lets it apply patches, --exec lets it run tests, builds and lints — and neither is on by default. --agent picks which harness CLI drives the loop, so this is not tied to one vendor. --transcript prints every turn, which is the way to see what the model actually asked for.