The newest Sidekick is the tool you’ll be reaching for when you want to reverse binaries quickly and accurately.

It is designed to help you search any binary, understand what it does, and extract what matters.

You ask a question. You get an answer backed by evidence. You verify the key points and move on.

There are other tools for applying AI to reverse engineering, but Sidekick is purpose-built for binary analysis. It’s already integrated, verifying, improving on hard targets. You get to focus on results, not the tooling.

A New Direction

The design philosophy behind Sidekick holds that the role of reverse engineers is to mediate between AI and the security outcomes that humans value. And, it answers the question (both today and looking forward): if/when AI excels at reversing, why humans?

In our sister field, software engineering, developers are still present, but their responsibility has shifted from coding to mediating between agents and outcomes that matter. They are tasked with keeping up with whatever the evolving definitions of success are.

So it is with reverse engineers. There will be less drilling down into disassembly or IL and much more directing. Sidekick is designed for a new mode of reverse engineering where:

  1. You ask questions at the altitude you care most about
  2. AI delivers hypotheses backed by evidence, fully transparent
  3. Redirection is fast — “no, let’s look at this instead”
  4. Drill-down is fast — one click from claim to IL to bytes
  5. Work is preserved for all stakeholders

Sidekick is designed to help you do your new job, fast, today.

What’s New in this Release?

Sidekick 26.0 is a redesign, not an iteration.

Agentic architecture

  • Specialist modes — Research, Transform, Repair Analysis, Modeling, Debugger, Automation, each with its own discipline.
  • Validation agent — cross-checks findings against IL, assembly, and memory traces. Catches type misrecovery, calling convention errors, and untested conditions; repairs decompilation and reopens the investigation.
  • Behavioral Model — persistent description of a binary’s behavior, carried across sessions and stored outside the BNDB.
  • Notebook — tracks goals, tasks, and outcomes (findings, artifacts, blockers) with explicit status: draft, verified, rejected.
  • Skills — reusable playbooks for situations like “hunt for use-after-free” or “recover types in stripped Go.” Load based on context.

Revamped UI

  • Inspectable chat tree — every subagent conversation is open for inspection. Nothing Sidekick does happens out of view.
  • Code Maps — auto-built call graphs and manual JSON diagrams. Live index entries render inside function nodes.
  • Indexes — BNQL, Script, or Manual. Scope across all binaries in a project. Metadata fields surface as sortable columns.
  • Notebook — Operational, Research, and Learning entries, with outcomes anchored to addresses.
  • Suggest menu — discrete operations for Repairs (function boundaries, signatures, indirect calls), Types, Names, and Comments.

Sidekick API

  • Completion Pools — map logical names (smart, fast) to one or more model endpoints with automatic failover.
  • Model templates — cloud Sidekick service, Ollama and vLLM for local models, direct Anthropic and OpenAI access.
  • Headless sidekick() — context manager opening the full workspace (chat, notebook, indexes, BNQL) to standalone Python.

Project support

We’ve extended Sidekick beyond a single binary so it can work across multiple binaries by leveraging Binary Ninja’s Projects feature.

  • Workspace — unifies chat, indexes, notebook, and semantic search across every binary in a Binary Ninja project.
  • Cross-binary BNQL — queries span all open binaries with attribution.
  • Project-scoped AI — investigates across the entire project, not one file at a time.

Also new

  • Semantic search — local vector database; the concept() operator works in BNQL, chat, and scripts.
  • MCP support — extend Sidekick with external tools via standard Model Context Protocol servers.
  • Debugger specialist mode — breakpoint-driven runtime observation with resumable threads.
  • Tool permissions — hierarchical tag-based controls scoped per-chat, per-workspace, or globally.
  • Transaction log — every database change is a transaction, with per-step undo.

Let’s Go

Experience Sidekick. Sign up for a free trial. Try it on something ugly. Share your feedback. (We’d love to hear from you!)