v0.84.8 · Kimi K2.6 default · DeepSeek V4 recommended

The terminal-native
AI coding agent
that ships your code,
not your data.

Open-source CLI for AI coding. Talks to Kimi, GLM, MiniMax, DeepSeek, and Ollama directly. No proxies, no telemetry, no $20/mo. 54 tools, 5 parallel agents, prompt caching.

Read the docs
~120K LOC · 100% Go 5 providers 54 built-in tools 5 parallel agents MIT license
~/work/payments-api · gokin
kimi-for-coding · 262K ctx $0.0243
connected
kimi · k2.6
5 agents · 2 active
tokens 12,418 / 262K
cache hit 94%
$ 0.0243
Why gokin

A coding agent without
the middleman.

Most AI coding tools route your code through their own servers for telemetry, analytics, and key management. Gokin doesn't. Every API call goes from your machine to the provider you chose. The code is open, you can check.

Feature
gokin
Claude Code
Cursor
Price
Free → pay-per-use
$20+/mo
$20+/mo
Providers
5 (Kimi · GLM · MiniMax · DeepSeek · Ollama)
1 (Claude)
Multi
Works offline
✓ via Ollama
no
no
Built-in tools
54
~30
~30
Parallel agents
5 (shared memory)
basic
no
Direct API calls
✓ zero proxies
routes through Cursor
Secret redaction
24 patterns, every tool result
basic
basic
Self-host / fork
✓ MIT
no
no
Source
Open · 100% Go
Closed
Closed
Features

What it can do.

Five agents. One shared brain.

Explore reads files in parallel. General edits them. Bash runs them. They share memory and pass tasks between each other. API retries, provider failover, and live streaming come built in. You see the work as it happens.

5 parallel 25 max turns worktree isolation auto failover

Extended thinking, kept across turns.

Multi-turn reasoning for Kimi K2.6, GLM, and DeepSeek thinking modes. Reasoning blocks carry their signatures across turns and tool calls. No "reasoning_content missing" errors when the agent picks up where it left off.

54 tools, classified.

Read-only tools run in parallel; write tools serialize. Slow tools get adaptive timeouts (p95 × 5). Failing tools downgrade their group to sequential.

MCP, with permissions.

Bring your own tools via Model Context Protocol. Each server gets a trust tier (low, medium, high) and every tool call inherits it.

$

Per-token cost, in your status bar.

Live spend with model-specific pricing. Explicit cache_control breakpoints cut input cost on Kimi, MiniMax, and DeepSeek by up to 90%. Gokin shows you when a break happens and how efficient the prefix was.

cache hit 94% retry 10× exp backoff circuit breaker 5/45s sandbox on
Multi-agent

Specialized agents.
Shared memory.

Up to five agents work in parallel, each one handling a slice of the job. They share a working memory, decompose tasks on their own, and stream output back as it happens. If one provider throttles, gokin moves to the next without dropping the conversation.

exploreread-only · runs in parallel · grep + read + glob
generalfull tool access · writes, edits, refactors
bashruns commands · sandbox · safe env whitelist
plandecomposes the task · todo · web research
guidedocs & how-to · web fetch + search
read-only
explore
write
general
orchestrator
coordinator
execute
bash
plan
planner
Providers

Five providers.
Pick one. Switch any time.

Four cloud providers, all daily-driver tier, tested every release. Ollama runs offline. Switch with /provider <name>.

★ DEFAULT
Kimi
kimi-for-coding · K2.6
262K context thinking + tool use coding-tuned
★ RECOMMENDED
DeepSeek
deepseek-v4-pro · v4-flash
1M context ~20× cheaper than Opus prompt caching on
★ BUDGET
GLM
glm-5.1 · glm-5 · glm-4.7
~$3/month coding plan z.ai endpoint
★ AGENTIC
MiniMax
M2.7 · M2.5 · highspeed
200K context strong agentic pay-per-use
⊙ OFFLINE
Ollama
any local model
0 network calls $0 cost fully airgapped
Security & Privacy

Your code never visits a third party.

Some CLIs route requests through their own Vercel Edge gateway for analytics. Gokin doesn't. Every byte leaves your machine over TLS, lands at the provider's endpoint, and comes back. The source is open, you can check.

  • 24 redaction patterns. AWS keys, GitHub tokens, JWTs, DB URIs, PEM keys. All stripped before reaching the model.
  • Sandbox mode. Bash runs in an isolated namespace, ~35 env vars whitelisted, API keys never leak to subprocesses.
  • 50+ blocked patterns. Fork bombs, reverse shells, rm -rf /, credential theft, env injection.
  • SSRF and path safety. Private IPs blocked, symlinks resolved, traversal prevented, TOCTOU checked.
  • 3-level permissions. Low auto-runs, medium asks once, high asks every time.
  • Keys stay local. Masked in UI, never in conversation history, never in tool results.
┌──────────┐         ┌──────────────────┐
│  gokin   │ ──TLS──▶│  Provider API    │
│  (local) │         │  (Kimi / GLM ...) │
│          │ ◀──TLS──│                  │
└──────────┘         └──────────────────┘

       no middle servers
       no Vercel
       no telemetry
       no analytics proxy

      your API key
      your code
      your conversation
      ────────────────────
      direct, every time.
Get started

One line.
Then start typing.

Requires Go 1.25+ if you're building from source. Otherwise the installer pulls a release binary for your platform.

macOS · arm64 / x86_64 Linux · arm64 / x86_64 Windows · WSL
1 · Setup
$ gokin --setup
Interactive picker for provider + key.
2 · Run
$ gokin
Launches the TUI in your repo.
3 · Ship
> add user auth, run tests, open PR
Gokin picks the right agent and reports back.
copied to clipboard