01 / Real tools
The full workbench, handed to the model.
Helix exposes its workbench to any compatible model over MCP: binary analysis and decompilation, a non-destructive decompiler overlay, patching and re-signing, wireless and network captures, findings, and more. Every capability below is a real tool the model can call, running the same surface the desktop uses.
Open, create, and switch between analysis projects, with live engine and GUI status.
Import binaries into a project and manage their storage and architecture.
Run a staged pipeline over a binary, from quick triage to full decompilation, or trigger individual capabilities on demand.
Turn native, .NET, DEX/APK, and JVM code into readable pseudocode and instructions.
Layer non-destructive renames, types, signatures, comments, and labels over the decompiler view.
Demangle and resolve symbols across binaries and trace string and data references.
Read Mach-O code signing, entitlements, chained fixups, and Apple sandbox profiles.
Match functions between two binaries and diff their instructions, pseudocode, and control flow.
Apply binary patches, rebuild and re-sign packages, and run binaries under a timeout.
Browse the project tree, read files intelligently, and unpack nested archives and firmware.
Capture findings as rich notebooks with runnable code playgrounds.
Seed structured findings for wireless captures - Sub-GHz regions, NFC block ranges, and RFID fields.
Build and run repeatable analysis workflows with their own data.
Import, browse, replay, and sniff HTTP traffic inside a project.
Import and dissect pcap captures with layered, Wireshark-style protocol detail.
Snapshot discovered network services and send audited TCP/UDP payloads.
Scan, capture, fingerprint, and interact with Bluetooth Low Energy devices and GATT.
Drive the paired phone companion for remote BLE actions and file transfers.
Decode Bluetooth Classic devices and phone-sourced MFi accessories.
Decode captured NFC tags, NDEF records, and APDU logs.
Decode RFID, Sub-GHz, IR, and iButton captures and their raw timings.
List and classify captured barcode and 2D-code payloads.
Merge every wireless and network domain into one unified capture timeline.
Import Flipper Zero SD captures and read from connected devices.
Detect and install the external toolchains that analysis depends on.
Open and drive real PTY shell sessions inside the workbench.
02 / Findings
Conclusions write into the project, not a chat window.
When the model concludes something, it calls helix_findings_create and writes a first-class finding into the project tree. The finding carries a node in the tree, .mdx content, base64 artefacts, and structured domain blocks that tie a conclusion to the exact captured record it came from.
The finding playground runs for real: python, node, bash, cpp, rust, and asm. stdout, stderr, and the folder diff come back on every run.
- First-class findings
The model writes named findings straight into the project tree, not a chat transcript.
- Rich notebook bodies
Each finding stores a structured document alongside a JSON sidecar.
- Structured wireless findings
Purpose-built findings for Sub-GHz regions, NFC block ranges, and RFID fields.
- Runnable playgrounds
Findings embed code blocks the model executes - python, node, bash, cpp, rust, asm - and reads back stdout, stderr, and the folder diff.
03 / Auditability
Auditable, replayable, consent-gated.
Driving a workbench from a model only earns trust if every move is on the record and the dangerous ones ask first.
Consent before any radio or network write
helix_ble_host_probe_gatt, helix_ble_host_gatt_write, and helix_network_send_payload all require consented=true in the call. If the flag is absent the tool refuses and returns an error before anything is sent. The response is audited to the project.
Analysis is staged and polled
There is no single one-shot full-analysis call. The model runs helix_analysis_run_stage to start a specific stage, then polls helix_analysis_list_stages until the stage reaches Done or Failed. Every intermediate state is visible. helix_analysis_ensure_stages with a named profile queues only the missing stages.
Results persisted on every path
helix_findings_create writes a typed finding into the project. That record is persisted the same way whether the work ran over the MCP path or through the app itself. The model can retrieve it after the fact.
Local by design, no cloud orchestration
Helix runs as a local process on the same machine as the project. There is no remote or cloud orchestration layer involved. Everything the model touches lives on that machine, and helix_ping confirms the engine is reachable before any GUI-routed call.
04 / Full surface
Every capability domain.
Every capability domain in the Helix MCP server, mapped as a single structured index of the complete surface.