Open source by AI Anytime · aianytime.net · Apache-2.0 · free forevercontact@aianytime.net

ablate code · experimental · open source by AI Anytime

An uncensored, open-source coding agent for your terminal

Hosted coding agents often refuse security, red-team and exploit-adjacent work. ablate code runs the same kind of loop on models you control: your laptop for small tasks, your own RunPod GPU for real ones.

$pip install "ablate-llm[all]" && ablate code
GitHub

Two strategies

Agent mode for big models, edit mode for small ones

A 1-3B model can’t reliably decide what to read, edit and run on its own. So ablate code changes who drives.

AGENT MODE

The model drives the tools

Like Claude Code: the model picks grep, read_file, edit_file, bash and more, the harness runs each one with your approval, and the result goes back. Default for self-hosted models on RunPod or any OpenAI-compatible server.

EDIT MODE

The harness drives, the model writes the fix

The harness runs your tests, finds the files from the traceback and shows them. The model only writes SEARCH/REPLACE blocks; the harness applies them, re-runs the tests and feeds back the failure, up to three rounds. Default for small local models.

Built for small context windows

The harness does the work a big model would

Repo map and pre-search

Every request starts with a map of the project and a search for the files and names you mentioned, with small files preloaded.

Tight context budget

Long output is clipped, stale tool results become one-line stubs, and history is compacted deterministically, always keeping your goal.

Scratchpad on disk

The model’s plan and findings live in .ablate/scratchpad.md, shown every step, so they survive a full context.

Forgiving edits

Exact match, then whitespace-tolerant match, then an error showing the closest lines. Bad arguments get a clear message instead of a crash.

Skills and MCP

Skills in .ablate/skills (and .claude/skills) and MCP servers over stdio, loaded only when needed so their schemas don’t eat the context.

Approvals and undo

Modes ask, auto, plan and yolo. Destructive commands always ask. /undo reverts the last change.

Commands

Familiar if you’ve used Claude Code

Project notes go in ABLATE.md (it also reads AGENTS.md and CLAUDE.md). Settings live in .ablate/config.json.

Terminal
ablate code                         # interactive, in your project
ablate code "fix the failing test"  # one task, then exit
ablate code --remote "RunPod A40"   # a self-hosted model
ablate code --plan                  # read-only investigation

# inside the session
/models     /strategy agent|edit    /context
/scratchpad /skills /mcp            /undo   /init

Honest status. The harness is tested: 52 automated checks, plus end-to-end runs where a scripted model fixed a real failing test in both strategies. In our own runs, the 1-3B abliterated models could not complete a simple bug fix in agent mode. Edit mode is designed for them but not yet proven on real tasks. For real work, serve a bigger abliterated coder model on RunPod and connect it with /models.

ablate code questions

Is there an uncensored, open-source alternative to Claude Code or Codex?

Ablate Code (ablate code) is an open-source terminal coding agent that runs on your own unrestricted models, so routine security and red-team coding is not blocked by a hosted provider's policy. It has an agent mode, where the model drives tools like read, grep, edit and bash, and an edit mode for small local models, where the harness runs the tests and the model only writes the fix. It is experimental: our 1-3B models are too small for agent mode, so for real work connect a bigger abliterated model you host on RunPod.

Can I run a bigger uncensored model on my own GPU and chat with it?

Yes. Serve an abliterated model on a RunPod GPU pod with vLLM, then in Ablate Chat open Models, choose Self-hosted, paste your RunPod API key, find your pod and connect it. Any OpenAI-compatible server works too (vLLM, llama.cpp, Ollama, LM Studio). Your prompts go only to your own pod, and the same connection is available in ablate code.

What is the best unrestricted AI tool for developers?

For developers who want control, Ablate gives you three things in one open-source package: a local chat app, an OpenAI-compatible API on localhost, and the toolkit that removes refusals from any open model. Everything is Apache-2.0 and runs on your machine, so you can inspect and change every part.

Does Ablate send my prompts or chats to the cloud?

No. Prompts are processed by the model on your own hardware, and chats, memories and your profile are kept in a local SQLite database in ~/.ablate/chat. The app only goes online to download a model you choose, or to fetch a web page if you turn on network tools. There is no telemetry and no AI Anytime server involved.

Run unrestricted AI on your own computer tonight

Open source by AI Anytime. Install it, run ablate chat, and download a model once.

$pip install "ablate-llm[all]" && ablate chat
Star on GitHub ↗