The agent

  ✓ 1 skill greet
  ✓ mcp arcflare 29 tools
  ✓ mcp blender 67 tools
  prefix 1064 tok · window 256K · 96 mcp tools indexed (schemas on demand)

❯ widget.js has a bug in add(). Fix it and verify with node.

  ⚙ read_file widget.js
  ⚙ edit_file widget.js
  ⚙ bash node -e "console.log(require('./widget.js').add(2,3))"
  └ 5

  Fixed. The bug was `a - b` instead of `a + b`.

In a session: /rc remote control · /update · /stats speed and tokens · /tools · /skills · /bye.

The prompt is append-only. The system message and the tool array stay fixed for the whole session. MCP tools and skill bodies arrive as tool results, never spliced into the prefix. Editing the prefix throws the cache away, and at ~250 tok/s prefill a 20k-token session costs about 80 seconds to re-read.

When the window fills, compaction trims the oldest tool results first, then drops whole early turns. It never touches the system prompt or recent turns. A dropped tool call takes its results with it, because a result with nothing it answers is context the model can't interpret.

Built in: read_file, write_file, edit_file, list_dir, glob, grep, bash, plus skill_load, tool_search and mcp_call. Edits fail loudly on an ambiguous or missing match rather than guessing. Shell commands go through an approval prompt and a refusal list for unrecoverable ones.

Servers come from .arcflare/mcp.json or .mcp.json in the working directory, then ~/.arcflare/mcp.json or ~/.claude.json, and connect over stdio or HTTP. Tools are namespaced server__tool.

The model sees a one-line index, calls tool_search to pull the schemas it needs, and runs one with mcp_call. For 120 tools that is 1,072 tokens instead of 7,573. And because the tool array never changes, the cache survives every turn.

A missed search costs two turns, which on a local model means seconds. So search ranks whole words in a tool's name above the same word in its description, and results are written as compact signatures, not raw JSON Schema.

## blender__screenshot_render
Run a real render and return the image. engine: 'EEVEE' (seconds) or 'CYCLES'
(much slower, photoreal).
(engine?: string="EEVEE", samples?: integer=64, resolution?: integer[], camera?: string)

Hosted servers that use OAuth: arcflare mcp login [server] signs in, and arcflare mcp logout forgets the token.

A skill is a folder with a SKILL.md and YAML frontmatter. Discovery reads only the first 4 KB of each, so the index costs ~15 tokens per skill. The body (often 1,000+ tokens) loads only when the model asks for it. Skills are read from:

• ~/.arcflare/skills — yours, everywhere
• .arcflare/skills — this project
• .claude/skills — existing Claude Code skills work as-is

Put an AGENTS.md in your project (or ARCFLARE.md; a CLAUDE.mdis read if neither exists) with the build and test commands and anything the agent should know. It's read at the start of every session, from the folder and the git root. /init asks the model to write a first one for you.

Undo covers files changed by the agent's own write and edit tools. Changes made by shell commands aren't tracked, and /undo says so.