ByteAsk

ByteAsk is a terminal coding agent for C and C++ that joins the debugger you already have open, then builds and tests the change before you review the diff.

The debugger is already attached

Linux lets one debugger hold a running program. If gdb is already stopped on the bug, a second debugger is refused. ByteAsk does not open that second debugger. You start a small bridge inside the gdb session you already have, and ByteAsk reads the live state through it, then waits before it changes anything.

The homepage walks through a deadlock in a program called orderd. gdb is stopped in OrderBook::cancel at order_book.cpp:214. An agent told to look from outside runs gdb -p and gets "process is already traced" plus "ptrace: Operation not permitted." Through the bridge, ByteAsk prints the backtrace, sees the risk thread holding risk_mtx_, and proposes a scoped_lock on book_mtx_ and risk_mtx_. The proposal sits until you approve it. The breakpoints you already set stay set.

• The homepage names Claude Code and Codex as the agents that get locked out, because the kernel allows one tracer per process.
• You can see the lockout without installing ByteAsk: while your gdb holds a process, run gdb -p in another shell.
• The tool that joins your session is debug_attach. Batch gdb, a stateful debugger session, and rr replay are separate tools.

The diff has already been run

The claim is that the diff you review has already been compiled with your flags, passed through sanitizers, and run against your tests. The homepage shows one worked example and says the text is what the tools printed.

• g++ 12.2 with AddressSanitizer, captured on 2026-09-03, reports a heap-buffer-overflow in OrderBook::cancel_all at demo/outputs/order_book.cpp:31. ByteAsk says that function walks a level array one step too far, and that this is the one bug in the file.
• The same demo tree's ctest run reports 3 of 3 tests passed: book_ops, risk_engine, and wire_codec.
• That trace is ByteAsk's own capture from the homepage, dated 2026-09-03. It is not an independent rerun.

The docs add a limit the homepage does not put next to that story. Shell commands are read-only by default, so the agent cannot compile or write files through the shell unless you pass workspace-write. The built-in C and C++ tools, including compile, sanitize, and debug, sit outside that sandbox and still run. On Linux the shell sandbox is bubblewrap, installed by the install script, and it falls back to full access if bubblewrap is missing. macOS uses Seatbelt. The default approval mode is on-request: read-only steps proceed, and anything that changes state is proposed first.

A quoted passage, or an explicit miss

For the facts a model is worst at inventing, ByteAsk searches a fixed library and returns the passage, or it says there is no confident match. Register offsets, MISRA rule numbers, and wire-protocol fields are the examples they give. The lookup sends the search query. Your code and prompts are not part of it.

• autosar_ref covers AUTOSAR C++14, MISRA C++:2008, MISRA C:2012, the C++ Core Guidelines, ISO C++ (N4950), and C11 (N1570).
• mcu_ref covers the STM32F4 reference manual RM0090, Cortex-M manuals, ESP32, Modbus, CAN, MQTT, and linker scripts.
• fpga_ref covers Xilinx UltraScale and 7-series, Vivado, Vitis HLS, and AMBA AXI.
• hft_ref covers Nasdaq ITCH and OUCH, NYSE Pillar, Cboe PITCH and BOE, FIX, and IEEE 1588 PTP.
• grid_ref covers IEEE 1547 and 2030.5, SunSpec, ENA G98 and G99, and AS/NZS 4777.2.
• scpi_ref and x86_ref cover instrument manuals and the x86-64 Software Developer's Manual.
• cpp_intrinsic searches a 7,126-entry Intel intrinsics list instead of recalling an intrinsic from memory.

There is a separate open-source server, ByteAsk-Embedded-MCP, under the MIT license. Its README says the repo ships a small sample backend so it runs out of the box, and that the licensed document set stays on the hosted endpoint. That repository is the lookup server, not the coding agent.

Whose model, and who pays

You can use ByteAsk's sign-in, your own API key, or a server you run. The three paths are not the same bill.

• Managed: an emailed sign-in link, no API key. The catalog they publish lists gpt-5.4 as the default, plus gpt-5.5, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, an auto router, claude-opus-4-8, claude-sonnet-5, claude-sonnet-4-6, claude-fable-5, gemini-3.5-flash, and gemini-3.1-pro-preview. That traffic goes through ByteAsk.
• Your key: stored on your machine at ~/.byteask/byok-config.json with mode 0600. Traffic goes to OpenAI, Anthropic, or Gemini and does not pass through ByteAsk. On the free plan this full access lasts 14 days. On Pro and above, your own keys are unlimited.
• Your server: vLLM, Ollama, llama.cpp, LM Studio, TGI, SGLang, or a LiteLLM proxy, added with byteask models add. Self-hosted models are not metered, and those prompts do not go to ByteAsk's servers.

On the free plan, ByteAsk-managed use is about 25 agent tasks a month. Debugging, memory and undefined-behavior checks, and performance tools are the limited ones. Build and test, code quality, compiler insight, and the document lookups are listed as unlimited.

What the pricing page prints

• Free is $0.
• Pro is $20 a month, or $16 a month if you pay $192 for the year. It includes $20 of managed usage a month, frontier models, priority runs, and email support.
• Max is $100 a month, with 5 times the managed usage of Pro, plus a fast mode and early access.
• Ultra is $200 a month, with 10 times the managed usage of Pro, and faster support.
• A company with more than 10 people needs Team or Enterprise to use its own models. The individual price list does not print the Team rates.

Inside the app, /usage shows two rolling windows, 5 hours and 7 days, as percentages rather than dollars. The dollar figure it does show is extra credit. A referral pays $10 of credit to each side when the other person installs with your code and signs up. The docs' plan table stops at Max. The pricing page also lists Ultra.

Two founders, and a toolchain they already knew

ByteAsk is two people. Anirudha Kulkarni is CEO. Pratyush Saini is CTO. Y Combinator's company page lists them as a Fall 2026 company in San Francisco, founded in 2026, with a team of 2. The homepage's own structured data calls the build version 0.1.11.

India Today reported on 25 September 2026 that they raised $1 million in pre-seed from Y Combinator, Entrepreneur First, and angels from quantitative trading firms, and that the company was founded in June 2026. The same piece says their previous company, LawSutra AI, was bought by Manupatra four months after it was founded, after more than 5,000 lawyers had used it. Kulkarni's YC bio says that product now serves more than 100,000 lawyers. His line in the India Today piece: "Building LawSutra taught us that models are only as useful as the context, verification and workflows around them."

• Kulkarni studied computer science at IIT Delhi. His YC bio says he worked on low-latency C++ at Optiver and Quantbox, and built a Python-to-C++ transpiler there.
• Saini, also IIT Delhi, was a core engineer on ThirdAI's NeuralDB and BOLT, then a founding engineer at Simbian, where the YC bio says he co-built a security agent that handles over a million incidents a year.
• India Today says the first customers they are aiming at are high-frequency trading, automotive, and embedded systems, and that a specialized C++ model is planned, not shipped.
• The same report says the funding is for research, GPUs, training data, and on-premises deployment. YC's blurb says the product can run fully on-prem.

Where it installs, and which editors are ready

• The docs list macOS on Apple Silicon, Linux on x86_64 and arm64, and Windows on x86_64 with no WSL. Install is a curl script, a PowerShell script, pip, uv, or npm install -g @byteask/cli.
• Sign-in is a magic link. Disposable addresses and .ac.in academic domains are refused.
• The homepage lists plugins for VS Code, Cursor, JetBrains, Neovim, Emacs, and Zed, plus Windsurf and VSCodium through Open VSX. Every plugin needs the byteask command on your PATH.
• The extensions repository marks Neovim, VS Code, and Emacs ready. Zed is built and not yet submitted to its registry. JetBrains is built and still waiting on a manual check. Helix and Sublime are marked planned. That table is more cautious than the homepage list. The extensions repo is Apache-2.0. The agent you install is the CLI, not that repo.
• Sessions are stored in a local SQLite database under ~/.byteask/. ByteAsk says prompts and code are dropped after the request and are not used to train models. The terminal UI is built with Ratatui and Crossterm.