For macOS · included in every plan

A terminal agentmade for open models.

Inferaft Code reads your repository, edits files, runs commands and checks its own work on Kimi K3, GLM 5.3, DeepSeek V4 and the rest of the catalog. It is included in every plan.

brew install inferaft/tap/inferaft-code

Get an API key See what it does

Apple Silicon Mac. Homebrew installs Node 24 and ripgrep.

~/acme — inferaftIllustrative session

❯ Fix the flaky retry test in billing and keep the fix small
  Thought for 4.2s, read 3 files, searched 1 pattern, ran 2 shell commands ›

● Agent(Map retry call sites) · Done (6 tool uses · 2.8s)

● Update(src/billing/retry.ts) · Updated with 6 additions and 2 removals

● Bash(pnpm test billing) · 42 passed in 3.1s

● The test waited on wall-clock time. It now advances a fake
  timer to the scheduled tick, so machine load no longer matters.

────────────────────────────────────────────────────────────────
❯ describe a task · / commands · @ files
────────────────────────────────────────────────────────────────
~/acme • main • CH91.2% • $0.0184 • 38.1k/262.1k (14.5%)    kimi-k3 • default

Everything a long coding session needs.

Plan, delegate, change, check and go back, with every step one line until you open it.

Every model, plus your own accounts

Switch with /model at any point. Connect a Claude or OpenCode Go account and its models join the list, grouped by provider. The status line shows context, cache hits and cost.

Sub-agents, side by side or in the background

Read-only explorers run four at a time, a background agent reports back when it is done, and your own agents live in .inferaft/agents.

A sandbox around every command

Shell commands run in a macOS sandbox: writes stay in the workspace and the network stays off. Leaving the sandbox always asks.

Worktrees built in

inferaft -w starts a session in its own git worktree, and a sub-agent can take one too, so parallel changes never collide.

Memory across sessions

Preferences, corrections and project context are kept as Markdown under ~/.inferaft-code. /memory shows, edits and deletes them.

Context that stays cached

A stable prompt layout keeps the cache warm, old tool output is trimmed first, and earlier turns are summarized when the window fills. /context shows every part.

Rewind with esc esc

Every prompt is a checkpoint. Go back to any of them and restore the code, the conversation or both.

Permissions you can read

Ask first, accept edits, plan or bypass, plus allow and deny rules such as Bash(npm run *). A deny rule always wins.

Artifacts you can open

Walkthroughs, dashboards and checklists are published as private pages on inferaft.com that update in place and can be shared.

Local by default

The key lives in macOS Keychain and sessions stay on your disk. No telemetry, no transcript upload, and project files never start servers or hooks on their own.

Your setup comes with you.

Inferaft Code reads the same files and formats as the other agents you use, so nothing needs rewriting.

AGENTS.md · CLAUDE.md

Read from the repository root down to where you work. A subdirectory's file loads the first time the agent touches it, and @imports work.

MCP

The shared mcpServers format over stdio, HTTP or SSE. A repository's .mcp.json starts only after you trust it.

Agent Skills

SKILL.md folders from ~/.inferaft-code/skills and ~/.agents/skills, so skills installed for other agents work here too.

~/.inferaft-code

One folder for settings, MCP servers, skills and commands. What you pick in a session becomes the next default.

# MCP servers and skills
$ inferaft mcp add --transport http notion https://mcp.notion.com/mcp
$ inferaft skills add anthropics/skills --skill skill-creator
$ inferaft mcp list

# ~/.inferaft-code/settings.json
{
  "model": "kimi-k3",
  "permissions": {
    "allow": ["Bash(npm run *)", "Edit(src/**)"],
    "deny": ["Bash(git push *)", "Read(.env)"]
  }
}

Made for the terminal.

Fullscreen by default, fast to redraw, and every detail one key or click away.

esc escRewind to an earlier prompt
ctrl+oExpand thinking, commands and output
clickOpen one step at a time
dragSelect text, copied when you let go
shift+tabCycle the permission mode
/Every command, aliases included
@Attach a file from the workspace
!Run a shell command yourself

And in scripts.

One non-interactive turn with -p, as plain text, one JSON result or a stream of JSON events. Changes are denied unless you allow edits.

$ inferaft -p "summarize why CI failed"

$ inferaft -p "list the TODOs in src" \
    --output-format json

$ inferaft -p "update the changelog" \
    --permission-mode accept-edits

Questions, answered plainly.

Is Inferaft Code included in my plan?

Yes. It uses the same API key and the same plan allowance as any other client. There is nothing extra to buy.

Can I use it without an Inferaft plan?

Yes. Connect a Claude or OpenCode Go account and the session starts with those models. Inferaft models and artifacts need an Inferaft key.

Which platforms are supported?

Apple Silicon Macs through Homebrew. On other systems, connect Claude Code, Codex, OpenCode or Cline to Inferaft instead; the docs have a block for each.

How do updates work?

It checks the Homebrew cask in the background at most every 20 hours and shows one line when a new version is out. /update or inferaft update upgrades through Homebrew.

What leaves my machine?

Model requests send your prompt and the context the agent reads to the provider of the model you picked. Sessions, memory and settings stay in ~/.inferaft-code, and Inferaft does not retain prompt or response bodies.

Is it related to Claude Code or Codex?

No. It is Inferaft's own agent and is not affiliated with Anthropic or OpenAI. It reads the same AGENTS.md, MCP and skills formats, so your setup carries over.

Install it, paste a key, start a session.

brew install inferaft/tap/inferaft-code

Get an API key Compare plans