agentbelt 0.3.0: token optimization for local AI agents, with honest numbers

Lova Rama Krishna Pothagani — Sat, 14 Nov 2026 08:08:07 +0000

agentbelt gives every AI coding app (Claude Code, Codex, Cursor, Copilot, DeepSeek, LM Studio) the same safe, audited dev tools over MCP. Version 0.2.0 added context compaction for agents running on local models with small context windows (gpt-oss 20B, 16k context, on a MacBook Air). 0.3.0 adds a Goose adapter and a demo GIF.

I measured it instead of guessing

I ran the same tasks with compaction off and on, 3 runs each. Token counts are what Ollama itself reported.

Task	Compaction	Correct	Avg input tokens	Largest prompt
read 5 files	off	3/3	46,916	13,317
read 5 files	on	3/3	50,410	10,269
failing test log, twice	off	1/3	42,267	6,182
failing test log, twice	on	1/3	18,448	3,338

The largest prompt was lower in every run.
On the log-heavy task: 56% fewer input tokens, and every run finished instead of running out of steps.
On the file-reading task: same accuracy, but no average saving. One run re-read files and cost more.
My first version did worse. It shortened file contents too much, and the model re-read what it had lost.

Takeaway: for local agents, the biggest win is not repeating the same logs. Shrinking file contents can backfire.

Full method and raw results: docs/token-optimization.md

Try it

npx agentbelt setup

Then ask your AI app: "Use agentbelt's agent_run with compact: true to …"

⭐ If this is useful, a star on GitHub helps other developers find it. And if you run agents on local models, I'd love to hear what you've measured.

One Toolbelt for Every AI Coding Assistant: Why I Built agentbelt

Lova Rama Krishna Pothagani — Fri, 13 Nov 2026 06:45:02 +0000

Claude Code, Codex, Cursor, VS Code, DeepSeek and LM Studio, with the same safe tools, set up once and tested for real.

The problem: every AI app reinvents the same tools

I use several AI coding assistants. Each one can run commands, touch git and run my tests, and each one does it differently: different config files, different permission prompts, different ideas about what's safe. When I switched apps, I lost my workflows. And when an assistant runs git reset --hard or pushes to the wrong branch, "the model meant well" doesn't help much.

I wanted one set of developer tools that:

works the same in every AI app;
is safe even if the model gets confused or manipulated;
I could set up in a minute and trust after an update.

The idea: tools as a shared, audited layer

The Model Context Protocol (MCP) lets AI apps call tools provided by a separate program. agentbelt is that program. It provides:

git tools that read the repo and do work in isolated task branches (agentbelt/<id>). They can commit there but never push, and never run reset, rebase, clean or checkout.
testing tools that find and run the project's own checks (npm, pnpm, yarn, Cargo, Go, Swift, pytest).
workflows: small JSON recipes like repo-health or safe-commit, with approval steps.

Your AI app brings the model. agentbelt brings the tools. No API key, no second LLM.

Safety by design, not by prompt

Every action goes through a broker. Plugins never run commands or touch files directly. The broker checks each request against the plugin's declared permissions and a core policy: no shell, allowed subcommands only, paths that can't leave the project or touch .git.
Only you can approve. A workflow's approval step can only be approved in your terminal (agentbelt approve). There's no tool for it, so the model can't approve its own work.
Everything is logged to .agentbelt/audit.jsonl, with secrets removed.
Honest labels. Built-in tools are labeled policy-only: checked, but not OS-sandboxed. Third-party plugins only run inside a real OS sandbox (macOS sandbox-exec, Linux bubblewrap) with no network, and if there's no sandbox, they're refused.

Updates you can trust

npx agentbelt setup installs a pinned copy with install scripts turned off, then makes real tool calls against it before switching to it. agentbelt update does the same with a new version next to the old one; agentbelt rollback switches back instantly.

"Compatible" means tested

An app counts as supported only when it made real tool calls in a real session, and the calls show up in agentbelt's audit log. Testing that way found real problems:

DeepSeek Harness doesn't tell tool servers which project is open, and starts them in its own settings folder. agentbelt now refuses settings folders and asks for --workspace.
Cursor doesn't fill in ${workspaceFolder} in its global config. With a folder open, Cursor does share the project through MCP, so the fix was to stop relying on the placeholder.
LM Studio needs the MCP plugin turned on in each chat, and a model whose tool calls it can parse.

Today agentbelt is verified with Claude Code, Codex, Cursor, VS Code (Copilot agent mode), DeepSeek Harness and LM Studio.

Try it

npx agentbelt setup

Then ask your AI app: "Use agentbelt to run the checks and fix what fails."

Guide: https://lovark.github.io/agentbelt/
Code: https://github.com/LovaRK/agentbelt (Apache-2.0)

⭐ If agentbelt is useful to you, a star on GitHub helps other developers find it.

DEV Community: Lova Rama Krishna Pothagani

agentbelt 0.3.0: token optimization for local AI agents, with honest numbers

I measured it instead of guessing

Try it

One Toolbelt for Every AI Coding Assistant: Why I Built agentbelt

The problem: every AI app reinvents the same tools

The idea: tools as a shared, audited layer

Safety by design, not by prompt

Updates you can trust

"Compatible" means tested

Try it