Kimi-K2.6 · Cloudflare Workers AI

Claude Code alternative, on your own Cloudflare account.

An open-source terminal coding agent powered by Kimi K2.6, MCP tools, local memory, and direct Workers AI execution — no API middleman.

Install View source

Get Kimiflare updates

Occasional release notes, technical write-ups, and early experiments on building AI coding agents on Cloudflare.

No spam. Only meaningful ships.

zsh

kimiflare Ready when you are.

  › Explain this codebase
  › Find and fix a bug
  › Refactor a file

Type a message or /help for commands · ctrl-c to exit · shift+tab to cycle modes

› add a /health endpoint

────────────────────────────────────────

thinking… I'll need to read the server file first, then add the endpoint.

⠋ read(src/server.ts)

✓ read(src/server.ts)

⠋ edit src/server.ts

Permission requested

tool: edit

action: edit src/server.ts

@@ -42,6 +42,10 @@
  app.get('/', …)
+  app.get('/health', (_, res) => res.json({ ok: true }))
  …

› Allow once Allow for this session Deny

✓ edit src/server.ts

Done — added /health that returns { ok: true }.

────────────────────────────────────────

■ Update landing page terminal (12s · ↑ 2.3k tokens)

✓ Read current UI components

■ Update terminal simulation

☐ Create PR

[edit] k2.6 · medium · thinking…

in 2,847 (1,203 cached) · out 412 · ctx 12% · 0.00321

›

Features

What it does

Plan / Edit / Auto modes

Plan mode blocks all mutating tools for safe research. Edit mode prompts per call. Auto mode approves everything for trusted tasks.

Live task panel

For multi-step work, the agent publishes a task list with progress icons, elapsed time, and token deltas. Multi-step work feels managed.

Image understanding

Drop image paths (PNG, JPG, WebP, GIF, BMP up to 5 MB) into any prompt. The model sees them inline — perfect for UI reviews, diagrams, and screenshots.

14 terminal themes

dark, light, high-contrast, dracula, nord, one-dark, monokai, solarized, tokyo-night, gruvbox, catppuccin, rose-pine. Live preview with Ctrl+T.

Paste collapse

Large pastes collapse to [pasted N lines #id]. Full content still goes to the model — scrollback stays clean.

Type-ahead queue

Type your next prompt while the model is still working. Queued prompts fire in order. Ctrl-C aborts current and clears queue.

Smart context management

Compiled context extracts structured state and archives raw tool outputs as recallable artifacts. Auto-compaction kicks in at ~80% usage. The model never loses track of what it learned.

Streaming reasoning

Toggle the model's chain-of-thought with /reasoning or Ctrl-R. See how it thinks in real time.

Session persistence

Every turn is auto-saved. /resume lists past sessions with message counts in a paginated picker. Never lose your place.

Smart permissions

Bash session-allow is keyed by the first token (allow all git commands). Write/edit show a unified diff before you approve.

262K context window

Read entire modules, large configs, and full stack traces without the model losing track. Direct to Cloudflare — no middleman.

MCP server integration

Plug in external tools via the Model Context Protocol — local stdio servers or remote SSE endpoints. GitHub, Sentry, docs search, databases, and more.

Prefix cache optimization

Deterministic system prompts, stable JSON serialization, and session-affinity headers maximize Workers AI prefix-cache hits. Cached tokens are billed at a discount — cost drops as the conversation grows.

Explicit cross-session memory

The agent never surveils your conversation. Memories are stored only when you ask — via remember, recall, and forget tools — with SQLite + embeddings for durable, privacy-respecting retrieval across sessions.

Architecture

kimiflare Node.js TUI
↓ user msg → agent loop → runKimi()
↓ POST SSE to Workers AI
api.cloudflare.com @cf/moonshotai/kimi-k2.6
↑ tool result ← tool executor ← tool_calls
↑ permission modal for write / edit / bash

Tools

Core tools

Tool	Permission	Description
read	auto	Read a text file (≤ 2MB) with optional line range
write	prompt	Create or overwrite a file. Shows a diff before approval
edit	prompt	Replace an exact substring. Fails unless unique match
bash	prompt	Run a shell command. Session-allow keyed by first token
glob	auto	Match files by pattern, sorted by mtime
grep	auto	Regex search. Uses ripgrep if available
web_fetch	auto	Fetch a URL, convert HTML → markdown (≤ 100KB)
tasks_set	auto	Publish a live task list for multi-step work

Plus LSP intelligence (hover, go-to-definition, references, diagnostics), cross-session memory (remember / recall / forget), and MCP extensibility for plugging in external tool servers.