Yoke
An agentic CLI coding harness for DeepSeek models β Agentic coding on the BEAM.
is an agentic CLI coding harness for DeepSeek models β engineered on Elixir & Erlang/OTP for reliability, resumability, and fearless parallel agent work.
The Pitch
DeepSeek's models are brilliant. But a brilliant model in a fragile shell is just a smart idea waiting for a crash. Yoke is the harness that treats agentic coding like the serious, long-running engineering task it actually is β not a toy REPL, but a supervised, crash-recoverable system with the resilience Elixir and Erlang/OTP are famous for.
If you write code with DeepSeek's
deepseek-chat
,
deepseek-coder
, or
deepseek-reasoner
, Yoke is the missing layer between "a model that talks" and "a system that ships."
Why Yoke?
Crash containment that actually works
Most agent harnesses are a single process: one bad tool call and the whole session dies, along with your context, your reasoning, and your momentum.
Yoke is architecturally different. It separates the agent's
cognitive state
(the
Brain
, an isolated
GenServer
process) from
tool execution
(the
Hands
). When a tool call fails, OTP supervision contains the damage β the Brain survives, the session survives, and you keep going instead of restarting from zero.
Sessions that live to fight another day
Walk away from a session and come back to it exactly where you left it. Yoke checkpoints everything. Resume with
yoke -c
or
/resume
. No more re-explaining your whole codebase to a model that forgot everything overnight.
Hot-reload your tools, keep your context
Your tooling evolves as fast as your project does. With
/plugins reload
, Yoke lets you hot-reload tools
without
losing the session context. Iterate on your agent's capabilities the same way you iterate on your code.
Parallel subagents across the BEAM
Big tasks shouldn't run one step at a time. Yoke spawns sub-tasks as parallel subagent processes across BEAM worker threads β genuine concurrency for genuine productivity, powered by the actor model that's been running telecom-grade systems for decades.
Rules, references, and rigor built in
-
Scoped rule engine
(
/rules) keeps the agent honest to your project's conventions. -
Context-reference expansion
(
@filename,@https://...) pulls in exactly the files and docs the model needs β no more copy-pasting context. -
First-class Ragex integration
(
/ragex) puts code indexing and semantic search inside the harness. -
Native static-analysis hooks
for
oeditus_credo,propwise,credo, anddialyzervia/linter.
The Tech
| Language | Elixir |
| Runtime | Erlang/OTP |
| Models |
deepseek-chat
,
deepseek-coder
,
deepseek-reasoner
|
| Ecosystem | MCP, Ragex, oeditus credo, propwise | |
Who It's For
- DeepSeek power users who want a professional, resumable, and crash-safe agent environment.
- Teams shipping real code who need rules, linting, and quality gates wired into the agent loop β not bolted on afterward.
- Developers who love Elixir and want their AI tooling to share the same OTP philosophy as their product.
Bottom Line
Most agent harnesses ask, "How smart is the model?" Yoke asks the harder question: "How reliable is the system around it?" The answer β supervision, checkpointing, hot-reload, and true parallelism β is what separates a demo from a daily driver. Built on the BEAM. Built to last. Built for people who actually ship.
See the Usage section below to install it, or explore the Getting Started guide for complete onboarding, practice manifests, custom plugins, and workflow tuning.
Documentation
Browse all →Usage
Prerequisites
Elixir 1.19+, Erlang/OTP 27+, git, and on Linux,
build-essential
and
libgit2-dev
.
Install
curl -fsSL https://raw.githubusercontent.com/Oeditus/yoke/main/install.sh | bash
or from source:
git clone https://github.com/Oeditus/yoke.git
cd yoke
mix deps.get
mix yoke.install
Run
yoke