Quick start

Install the CLI and sign in once. Signing in opens your browser; GitHub or Google creates a free account if you don't have one, and no card is asked for.

Already installed? chi --version should print the current release. On Windows, or if you want to read the installer first, see Install.

In any repository you work in, from its root or any subdirectory, run:

One command does both halves of setup: it builds the index, and it connects the coding agents it finds on the machine. It is safe to run again: it asks only what is still open, which also makes it the way to connect an agent you install later.

Every question has a default that does the conservative thing. Steps that edit a file you track in git, or that would send anything off the machine, default to no. To see the steps before answering anything, chi init --list prints them and where each one stands, and asks nothing.

Indexing is the one step that takes real time, in proportion to the repository. It runs on your machine, because your code never leaves it; see Privacy and licensing.

chi init already connected the agents it found. Confirm the one you use:

chi agents check verifies the connection end to end: it asks the agent itself whether it can see the server, runs the exact command that was registered and makes a real tool call. A registration the agent cannot resolve fails here, not silently in your next session. To add an agent you installed later, run chi agents on <id>; the ids are on Coding agents.

Open your agent in the repository and ask something that would normally make it open several files. For example:

• Where is the retry policy applied, and what calls it?
• What breaks if I change the signature of this function?
• Walk me through how a request reaches the database.

It is working when the agent's tool calls come from Archis. In Claude Code they are named mcp__chi__chi_context, mcp__chi__get_files, mcp__chi__search and so on, where it would otherwise call Read and Grep. Other agents show the same tool names without the prefix.

Registering the tools puts them in front of the model; it does not make the model use them. That is why chi agents on also writes a short briefing into the instructions file your agent already reads (CLAUDE.md, AGENTS.md and so on), and for Claude Code installs a session hook, which works even when the model skips the file.

At the end of a task, chi session report shows what Archis served and what it saved, per tool and per turn.

It prices the same content two ways with a real tokenizer, full source against what Archis would serve, and reports the reduction. Pass a file, a directory, or nothing for the whole repository. Use --tokenizer cl100k if that is the one your model uses.

That is how much smaller Archis's view of your code is, not your bill: a real session also pays for the agent's answer and reasoning. What we measured in real sessions is on Check it yourself.

Optional, and useful the first time:

chi status is the command to remember: what the index holds, when it was last built, and how many files have changed since. --languages shows how deeply Archis understands each language in this repository. For the installation itself, run chi doctor.

search finds a symbol by name or signature and prints its id; symbol shows it, and --refs adds who calls it and what it calls. Add --impact for everything that transitively depends on it.

• Core concepts: what a symbol, a relation and a context level are.
• Coding agents: every supported agent and the tools Archis gives it.
• Tool output: the second saving, on what your tests and builds print.
• Using Archis on your company's code? Free covers repositories you own; an organization's repositories need a Team workspace, and its first two seats are free.