A coding agent that fits in 636 KB – no Node, no Python, no NPM

Developer descz launched Delta Agent, an autonomous terminal-based coding assistant distributed as a standalone C11 executable that weighs 636 KB on Windows and runs without external runtime dependencies such as Node.js, Python, npm, or pip. As detailed in github.com's report, the tool operates entirely inside console environments while consuming roughly 1.7 MB of private memory during execution. Users interact through natural language commands to inspect codebases, execute shell tasks, generate diffs, manage version control checkpoints, query external web endpoints, and retain session facts across runs using arbitrary API providers configured by the user.

This report examines the internal architecture, memory footprints, platform support constraints, local storage organization, and tool sets packaged into the single binary release. Delta Agent targets software engineers, terminal purists, and systems developers who require lightweight AI development tooling capable of operating across low-resource machines, locked-down enterprise environments, or resource-constrained continuous integration environments without installing sprawling package ecosystems.

A coding agent that fits in 636 KB – no Node: what actually changed

Delta Agent replaces multi-gigabyte language runtimes and complex interpreter toolchains with a compiled C11 executable that incorporates its own terminal user interface, networking engine, and file-parsing routines directly into native machine code. Rather than relying on heavyweight JavaScript engines, package trees, or multi-step environment preparations, the executable operates immediately upon delivery across x86_64 environments. On Windows, the binary installer measures 636 KB, requires approximately 1.7 MB of private working memory, and implements its own local HTTP setup server via setup.c to guide users through initial folder selections and path configuration without background server residue.

The native architecture fundamentally changes how autonomous agents handle turn-by-turn interactions and state management in terminal sessions. The application uses WinHTTP on Windows and libcurl on POSIX platforms within provider.c to process server-sent events line by line, enabling instant mid-stream request cancellations when users submit new instructions mid-turn. Instead of reprocessing full repositories on every prompt, Delta indexes codebase symbols, declarations, and imports once into cached project graph files, querying pre-indexed relationships via dedicated graph routines to keep frame rendering around 0.5 milliseconds per turn.

A coding agent that fits in 636 KB – no Node: how it works

The internal execution pipeline divides agent responsibilities strictly between local orchestration code and remote language model completions. Delta controls the operational loop while treating the remote model as a query engine, converting user input into streamed completions that trigger native tool operations across more than 40 discrete functions grouped into filesystem, shell, git, web, memory, and assertion families. Rendering handled in ui.c maps markdown structures into styled line documents through an internal cell grid utilizing OkLCH color tokens, while term.c interfaces with native raw console modes via ReadConsoleInput on Windows and standard termios on POSIX systems.

Repository awareness and workspace modifications are governed by local validation layers and automated rollback caches. The codegraph.c module stores structural relationship maps under the local user directory to resolve symbol cross-references without invoking recursive file reads. Every filesystem mutation preserves an undo snapshot to allow rollback operations via the /undo command, while version control tools generate status checkpoints before altering files. Context management rebuilds system instructions on every turn, incorporating active project graph states, indexed SKILL.md task workflows, and durable text facts retrieved from local JSON configuration records.

A coding agent that fits in 636 KB – no Node: why it matters now

Modern coding assistants routinely impose significant host overhead, demanding hundreds of megabytes in runtime dependencies, node_modules caches, and multiple virtual environments before executing a single completion. Delta Agent demonstrates that complete terminal interfaces, streaming network clients, diffing tools, and MCP protocol integration over HTTP or stdio can execute within minimal resource envelopes under 2 MB of RAM. This small footprint enables developers on legacy hardware, constrained virtual private servers, or strictly partitioned systems to operate full-capability autonomous development agents without altering system-level package manifests.

The system also challenges prevailing agent distribution patterns by isolating all runtime state, model configurations, and memory entries inside a single user directory located by default at .deltaCore. Operating without background daemon requirements, the executable ships with an embedded setup interface that presents a temporary local browser window during installation and shuts down cleanly once configuration finishes. Dynamic model switching and catalog refreshes sourced via models.dev give developers access to multi-provider inference pools without locking workflows into a single proprietary vendor desktop application or recurring subscription framework.

A coding agent that fits in 636 KB – no Node: who is affected

Software engineers working in resource-constrained environments, locked-down corporate workstations, and remote SSH terminal sessions experience immediate workflow differences. Developers without administrative privileges to install Node.js runtimes or Python wheels can download the self-contained executable, configure API keys via the /login command, and begin editing repositories immediately. Teams operating across mixed platforms must note disparity in test validation: while Windows x64 represents the reference platform with 283 end-to-end automated checks, Linux x86_64 has completed 32 core verification tests under WSL, leaving desktop-specific integration points such as clipboard managers and desktop entries unvalidated.

Organizations evaluating Delta Agent must also navigate distinct licensing and platform limitations during deployment. The software license permits source inspection, personal usage, and internal organizational operation while explicitly prohibiting commercial reselling or unlicensed cloning. For Apple hardware environments, developers cannot download pre-built binaries due to cross-compilation constraints across Mach-O targets; macOS 11+ users must compile the source locally via Xcode Command Line Tools through an included shell script, treating the platform implementation as an experimental starting point rather than a certified deployment binary.

A coding agent that fits in 636 KB – no Node: what to watch

Future releases will determine whether the developer can validate native desktop integrations across Linux and provide stable, pre-compiled macOS distributions. Because the author lacks Mac development hardware, community testing across Darwin architectures remains essential to confirm osascript automation, application bundle behavior, and dock integration. On Linux platforms, upcoming development passes must verify desktop shortcuts, GNOME favorites registration, and native clipboard utilities like wl-copy and xclip during active graphical display sessions to reach feature parity with the Windows reference implementation.

Adoption will also depend on how community members expand the built-in library of six default workflows covering code design, spreadsheets and PDFs, documentation generation, automation, verification, and systematic debugging. Observers should track updates to the native MCP implementation across stdio and HTTP transports, automated in-place binary swaps executed via delta update, and ongoing community discussions facilitated through the project's linked Telegram channel. As language models evolve, Delta Agent's capacity to handle reasoning effort levels and verbosity settings without expanding its 636 KB footprint will test the viability of pure C implementations.