Breaking

CodeBrain Released: LSP-Based Code Intelligence Cuts Agent Token Cost 63.9%

June 30, 2026 at 08:37 EDT

Software Dev & Coding
AI Agents
Open Source

Developer Tom Dörr has released CodeBrain, an open-source library that adds LSP-based code intelligence and dynamic planning to AI coding agents, reportedly cutting token cost by 63.9% in a full Terminal Bench 2.0 evaluation.

Feb 10, 2026 · CodeBrain (CodeBrain-1)

Open-source LSP intelligence cuts AI agent token cost by 63.9%

CodeBrain is a pure-Python library and MCP server that gives coding agents LSP-powered symbol search, diagnostics and dynamic planning — streamlining the plan-implement-verify loop on long-horizon tasks.

63.9%

Token cost reduction on full Terminal Bench 2.0

72.3%

Score on 47-task subset of Terminal Bench 2.0

LSP-powered agent tools, across 4 language families

Full 89-task evaluation · token cost

Claude-Opus-4.6 alone vs. paired with CodeBrain

$313.0

Opus-4.6
alone

$112.9

+ CodeBrain

Cost drops to roughly one-third — a $200 saving per full run.

How it works — the agent loop, tightened

Plan

Dynamic planning, adjusted during execution

→

Implement

LSP symbol search & impact checks

→

Verify

Diagnostics & validation, fewer iterations

Registers to Claude Code or OpenCode over stdio via MCP. Falls back to pyright, go vet or tsc where LSP is unavailable, and auto-detects subprojects in monorepos.

Praise

Improves code-search accuracy, shortens the edit-validate loop
Especially effective for large codebases and refactoring

Caution

Environment-dependent LSP server setup
Possible accuracy loss during fallback; few production cases so far

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.