Install AXLR

1. Download and verify

Every release archive contains axlr, axlr-tui, axlr-serve, VERSION, LICENSE, NOTICE, a short README.md and the third-party licence texts (THIRD_PARTY_LICENSES/), with a sibling .sha256. The current release is v0.10.0; all releases are on GitHub.

Linux and macOS (OS is linux or darwin, ARCH is amd64 or arm64):

V=v0.10.0; OS=linux; ARCH=amd64
base=https://github.com/underpass-ai/AXLR/releases/download/$V
curl -fsSLO "$base/axlr-$V-$OS-$ARCH.tar.gz"
curl -fsSLO "$base/axlr-$V-$OS-$ARCH.tar.gz.sha256"
sha256sum -c "axlr-$V-$OS-$ARCH.tar.gz.sha256"   # macOS: shasum -a 256 -c …
mkdir -p ~/.local/bin && tar xzf "axlr-$V-$OS-$ARCH.tar.gz" -C ~/.local/bin axlr axlr-tui axlr-serve
axlr-tui --version   # ~/.local/bin must be on PATH

Windows (PowerShell; ARCH is amd64 or arm64):

$V = 'v0.10.0'; $ARCH = 'amd64'
$base = "https://github.com/underpass-ai/AXLR/releases/download/$V"
Invoke-WebRequest "$base/axlr-$V-windows-$ARCH.zip" -OutFile "axlr-$V-windows-$ARCH.zip"
Invoke-WebRequest "$base/axlr-$V-windows-$ARCH.zip.sha256" -OutFile "axlr-$V-windows-$ARCH.zip.sha256"
$want = (Get-Content "axlr-$V-windows-$ARCH.zip.sha256").Split(' ')[0]
if ((Get-FileHash "axlr-$V-windows-$ARCH.zip" -Algorithm SHA256).Hash.ToLower() -ne $want) { throw 'checksum mismatch' }
Expand-Archive "axlr-$V-windows-$ARCH.zip" -DestinationPath "$HOME\axlr"
& "$HOME\axlr\axlr-tui.exe" --version

On Windows, use native executable paths and PowerShell environment syntax, and supply an absolute HOME or the XDG directories. You need a terminal of at least 50 × 15 cells.

2. Run the setup assistant

axlr-tui --root "$PWD"

The first launch in a terminal, with no settings.json yet and no --model, --session or -c, runs the setup assistant before the console. It asks two things at most and decides the rest; Enter takes the option marked *. axlr-tui setup runs it again at any time, for example to add a local model.

1. Model. With OPENROUTER_API_KEY set nothing is asked: the lightest recommended model becomes the default. Otherwise the assistant looks for OpenAI-compatible servers on this machine's usual ports (8080–8082, 8000, 11434 and 1234) and offers three routes:
   • OpenRouter: it asks for a key without echo, checks it against OpenRouter's model list, saves it in credentials.json beside settings.json, and recommends light models able to call tools.
   • A local server such as llama.cpp, vLLM, Ollama or LM Studio: pick a model it found or type a base URL. It reads the context window llama.cpp or vLLM report (for other servers, set context_tokens yourself) and writes a local_models entry.
   • Skip, and choose with /model later.
2. Jev, optional. TypeSafe's external judge: it checks final answers, answers the model's axlr_judge and settles each turn's memory, and that text goes to TypeSafe. With TYPESAFE_API_KEY set it is on; otherwise one hidden question, Enter skips.
3. KMP and MADE. On Linux (amd64, arm64) and Apple-silicon macOS they are installed without a question: the latest official releases are downloaded and verified, MADE's store and authorization are set up, the console's work grant and its pinned workflow definitions are published, and both engines are registered as MCP servers. On Windows and Intel macOS, connect them by hand with the KMP runbook and the MADE runbook.

Without a terminal (a script, a service) the assistant asks nothing: the key comes from the environment, and on the supported platforms the engines install themselves all the same.

A local model must return native tool calls: run llama.cpp with --jinja and a template that declares tools, and vLLM with --enable-auto-tool-choice and the parser that matches the model. Ollama truncates long prompts silently unless OLLAMA_CONTEXT_LENGTH is set. See local models.

The two engines

KMP is local-first agent memory: it keeps what was decided, the evidence it rested on and when it happened, never the transcript. It runs on your machine as one binary with an embedded SQLite store, with no account, service or API key, and it answers from stored evidence or says UNKNOWN instead of guessing. At every launch the console opens the checkout's own .kernel store inside a git repository, or a per-user store outside one; /memory save exports the project's memory to .kmp/memory.jsonl, which you may commit, and a new store imports it. ChronoLoom, the read-only local viewer inside the KMP binary, lets you move through that memory in time. Apache-2.0, pre-1.0: github.com/underpass-ai/kmp.

MADE coordinates: it runs no agents and calls no model. The console does each step's work, while MADE decides which step may be claimed next, enables a transition only when its guards hold, and records every accepted result in a hash-chained journal that verifies its own integrity and order, not the truth of the work. AXLR uses it locally, as one MCP process (made-mcp) over one SQLite file; MADE also runs as a gRPC service. Apache-2.0, pre-1.0: github.com/underpass-ai/made.

/update installs or updates both to their latest releases. The axlr-serve image pins its own engine versions.

3. First prompt

New sessions start in /normal. Press F1 for the controls and Ctrl+P for the action palette. After setup, autonomy is on: tools the work mode permits run without a card, except commands on the destructive list (deletions outside the workspace, git push, reset --hard, clean, rebase, branch deletion, network writes), which always ask. Reads never need a card. /approvals shows the saved choices.

Command	Does
/model	change the model
/review, /writer, /research	tailor the local tool policy to findings, documentation or evidence gathering
/debug, /delivery, /incident	run a built-in agent workflow on the next prompt
/plan	open the plans panel; /plan <goal> plans that goal in atomic tasks
/repair, /improve	work in a clone opened with axlr-tui --repair or axlr-tui --improve
/changes, /undo	review this session's file edits as diffs; put back the last turn's files (not what a command or MCP server did)
/tools	list, search, install and publish catalog tools
/mcp, /plugin	MCP connections and AXLR-managed packages
/update	install or update KMP and MADE (Linux, Apple-silicon macOS)

A turn makes at most 32 tool calls (turn_tool_calls, 8–256) before it pauses; send a message to go on. max_session_usd caps a session's spend and /usage shows tokens and cost. axlr-tui -c reopens the workspace's latest session and --session <id> a given one.

Commands run with your rights unless you turn on exec_sandbox: on Linux it runs every command under bubblewrap with the filesystem read-only except the workspace, optionally without network. It is off by default. Extend AXLR with any MCP server (stdio or Streamable HTTP) and with compatible OpenAI/Codex plugin packages, under the same approvals.

The console sends model requests to OpenRouter or to the local server, and to TypeSafe when Jev is on, and saves sessions under $XDG_STATE_HOME/axlr/sessions. Local tools act on the chosen workspace; exec and MCP servers use your account's permissions, so run AXLR in an account and workspace you trust. More in getting started and the configuration reference.

The other interfaces

• The JSON worker needs no model account; it offers read, write, edit, exec, search, list and the plugin calls: printf '%s\n' '{"protocol_version":1,"request_id":"hello-1","tool":"read","arguments":{"path":"README.md"}}' | axlr --root "$PWD" --profile trusted-local. See the worker contract.
• The service, axlr-serve, exposes HTTP, gRPC and MCP with mTLS. It needs remote KMP and MADE, runs normal sessions only (no work modes, workflows or plugin packages) and deploys on Kubernetes with the Helm chart as one replica, with mTLS ending in the pod. See the service API.
• The Go library embeds the executor, the OpenRouter client and the MCP client in your own host. See the library guide.

Build from source

With Go 1.26, from a full checkout of underpass-ai/AXLR:

GOWORK=off go test ./...
GOWORK=off go -C tui test ./...
GOWORK=off CGO_ENABLED=0 go build -trimpath -o /tmp/axlr ./cmd/axlr
GOWORK=off CGO_ENABLED=0 go -C tui build -trimpath -o /tmp/axlr-tui ./cmd/axlr-tui
/tmp/axlr-tui --root "$PWD"