Koder

aiXplain's agentic coding environment — your terminal, your tools, your models.

Install · First run · Using it · Features · Upgrading · Platforms · Configuration

Koder runs in your terminal, reads and edits the project you point it at, runs your tools, and works against the models in your aiXplain account.

• Run any model — you're not tied to one vendor. Claude Code speaks only to Anthropic and Codex only to OpenAI. Koder runs on the full model catalogue in your aiXplain account from a single API key, or on any provider you bring — Anthropic, OpenAI, Google, OpenRouter, local models, and more. Change models mid-session without changing tools.

• Bring your history with you. Coming from another tool? koder import pulls your past sessions in from opencode, Claude Code, and Cursor, so you pick up where you left off instead of starting cold.

• One agent, wherever you work. The same Koder runs as a fast terminal TUI, a headless server (koder serve) for scripts and CI, and a browser interface (koder web).

• Built to extend. Connect MCP servers for extra tools and data, run specialized subagents in parallel, and switch between agents with a single keypress.

• Yours to run. Open source under MIT, shipped as one signed-and-notarized binary that installs without root. Your API keys and session history stay on your machine.

curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash

Open a new terminal afterwards, then run koder from inside a project:

cd ~/your-project
koder

The installer drops a single binary in ~/.koder/bin and adds that directory to your shell's PATH. It does not need root, and it touches nothing else.

To install a specific version, or to skip the PATH edit:

curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash -s -- --version 0.3.1
curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash -s -- --no-modify-path

Koder asks for an aiXplain API key the first time it starts.

Sign in at app.aixplain.com, open Integrations → API keys, create a key and paste it in. That single key unlocks the model catalogue on your account. It is stored under your home directory, never in your project.

If you would rather use a different provider, choose Use a different provider on the same screen, or run koder providers later.

Start Koder in a project and describe what you want in plain language. It reads the files it needs, proposes edits, and runs commands when you let it.

cd ~/your-project
koder

koder run exits non-zero when a turn fails, so it composes with && and CI gates.

Run the full catalogue on your aiXplain account from one API key, or bring your own provider — Anthropic, OpenAI, Google, OpenRouter and more. Switch models mid-session without restarting.

Local models work out of the box. If Ollama is running, Koder finds it and offers exactly the models you have pulled — asking Ollama which of them can hold a conversation, call tools, accept images and how much context they have, rather than guessing. Nothing to configure; set OLLAMA_HOST if yours is not on the default port.

koder run --model ollama/qwen3.5:latest "explain this repo"

Switch agents with tab. build edits, plan investigates without touching files, and specialised agents can run in parallel as subagents. koder agent manages them.

• MCP servers — koder mcp connects external tools and data sources.
• Powers — MCP servers that stay dormant until a keyword in the conversation wakes them, so a big toolbox costs nothing until it is wanted.
• Workspace helpers — scripts the agent writes for itself become callable tools.
• Claude Code compatibility — reads Claude Code agents, hooks, plugin manifests and marketplaces.

• Specs (koder spec) — turn a request into requirements, a plan and tasks, then run the tasks in concurrent waves.
• Workflows (/workflows) — model-authored orchestration scripts you can pause, resume and stop by name.
• /loop — schedule a recurring prompt, list what is scheduled, and cancel one.

/memory, /dream and /flush manage what Koder remembers about a project. Corrections you make are recorded as lessons and re-applied in later sessions.

Drive a real browser: navigate, snapshot the accessibility tree, manage tabs and record sessions for replay. Opt-in — see KODER_EXPERIMENTAL_BROWSER.

Koder is built to be run on real repositories, so the limits are first-class rather than advisory:

Commands run inside an OS sandbox where one is available (seatbelt on macOS, bubblewrap on Linux), with sandbox.deny globs covering both shell commands and the file-writing tools. Administrators can pin policy that a project config cannot reopen.

koder upgrade       # move to the newest release
koder upgrade 0.3.1 # or a specific one
koder uninstall     # remove the binary and everything under ~/.koder

koder upgrade re-runs the installer for the version you asked for, so it follows the same path as a fresh install. Each release is kept in its own directory, so moving back to an earlier one is the same command.

On 0.3.0 or earlier? Those builds mis-detect how they were installed, so koder upgrade fails on them with Unknown method: unknown and background updates never run. Re-run the install command once to get onto a build that updates itself from then on — your configuration, credentials and session history are untouched:

curl -fsSL https://github.com/aixplain/koder/releases/latest/download/install | bash

macOS 12 or later. On Windows, download koder-windows-x64.tar.gz from the releases page and put koder.exe somewhere on your PATH; the shell installer above is for macOS and Linux.

macOS builds are signed with aiXplain's Developer ID certificate and notarized by Apple, so they run without a Gatekeeper prompt. You can confirm that yourself:

spctl --assess --type install -vv ~/.koder/bin/koder
# koder: accepted
# source=Notarized Developer ID
# origin=Developer ID Application: aiXplain, Inc. (U6RQVYMH8L)

To check an archive against the checksum on its release:

shasum -a 256 koder-darwin-arm64.tar.gz

The command was renamed to koder. The old names — aixplain-code, aixplain and aiXplain — still work and point at the same binary, so existing scripts and muscle memory keep working. Release archives are published under both names for the same reason.