3code Economical Coding Agent
Treats your token budget as a first-class constraint — 5× the work per token so subscriptions last and free tiers work.
Per-dimension scores and reasoning
The README mentions a loop guard, a private mode (-p restricting turns to zero-training providers), no telemetry, and local session storage, showing awareness of permissions and data flow. However, no sandbox implementation, permission manifest, confirmation prompt, or data-flow documentation exists in the repository to verify these claims; installation is a remote curl | sh with no auditable in-repo script; dependencies resolve by name from a public registry with no lockfile or hash pinning. These are assertions without support, so each scores 1. No malware, credential theft, or covert exfiltration was found, so the status is not blocked.
CI workflow comments show the author fixed a real failure-visibility bug (nimble swallowing non-zero testament exit) and added a watchdog plus timeout, which is genuine investment in failure surfacing. But there are no error-message samples, retry/degradation strategy, or dependency-availability checks in the repo, and README claims such as '40 known-good combos' and '1.6 MB single binary' have no supporting files. All three criteria score 1.
The README clearly covers macOS/Linux/Windows/Termux install paths, an NVIDIA free-tier quickstart, any OpenAI-compatible endpoint, and embedding as a Nim library, so audience and scenarios score 2. Capability boundaries (sandbox scope, tool permissions, autonomous-edit stop conditions) are only a one-line loop guard mention, trigger precision (when it acts vs. asks) is undocumented, and environment fit lists platforms but omits version/prerequisite detail, so those three score 1.
The MIT LICENSE file is complete and consistent with the README, scoring 3. The README is well structured and install notes cover three platforms plus a Windows manual alternative, scoring 2 each. There is no CHANGELOG, no versioning policy, no FAQ, and no known-limitations section; naming stability (CLI flags, .3log format, library API) carries no compatibility promise; maintenance responsibility rests on an individual trademark and blog, so those score 1.
The README presents a SWE-bench 10-task token comparison claiming 9% fewer tokens than pi and one more task resolved, with links to a blog and detail page, which is a usable value proposition. But the sample is only 10 tasks, self-run by the author, with no third-party reproduction; output usability (generated code quality, reversibility) has no evidence; and cost-benefit rests on unverified benchmarks. All three score 1.
Benchmark data points to external sites and a blog, with no raw results, scripts, or methodology in the repository, so it cannot be traced at source level. Marketing claims ('5x more work', '75% fewer tokens') coexist with the 88.3% figure in the table, and fact versus inference is not clearly separated. There is no cross-source corroboration beyond the README and CI comments. Each scores 1.
- Installation uses curl | sh and irm | iex, fetching and executing from a remote domain with no auditable in-repo installer; download and review the script before running.
- Dependencies resolve by name from a public Nim package registry with no lockfile or hash pinning, so build reproducibility and supply-chain safety cannot be confirmed from source.
- Token-savings and SWE-bench results are a self-run 10-task subset with no raw data or reproduction scripts in the repo; do not treat them as the sole basis for adoption.
- Sandboxing, permission boundaries, autonomous-edit stop conditions, data flow, and sensitive-data handling are only one-line README claims with no implementation or documentation backing.
- Publisher identity is unverified by the FollowAgents enterprise registry; maintenance responsibility and update path rest on an individual domain and blog, creating a single-point-of-failure risk.
What does this agent do, and when should you use it?
3code is a command-line coding agent whose core design goal is cost rather than raw capability, explicitly targeting more work per token so a subscription lasts longer and free-tier models stay usable. It works with any OpenAI-compatible endpoint, so free tiers, flat-rate coding plans, subscription logins and local servers are all valid backends. It ships as a single 1.6 MB executable with no runtime dependencies, runs on Linux x86-64/arm64, macOS universal, Windows and Termux on Android arm64, and has no daemon, no web UI and no telemetry. Internally it keeps context small through chunked mode (extract relevant context, discard stale material), aggressive caching, supersede-aware context compaction that elides stale reads, and a self-clearing plan/execute skill that resets context between phases. It is also usable as a Nim library: the same agent loop, tool calls, sandbox and session persistence, with the terminal replaced by return values and callbacks. It publishes a 10-task SWE-bench Verified subset comparison showing the lowest token use of five harnesses on the same provider and model.
As a CLI agent, 3code runs inside a project directory: it reads repository files, calls tools to edit them and execute commands, decides its next step from model responses, and produces code changes plus runnable results. All model traffic goes through an OpenAI-compatible endpoint, which can be a free tier, a flat-rate coding plan, a subscription login or a local server. To hold cost down it continuously extracts and discards context in chunked mode, reuses aggressive caching, performs supersede-aware compaction where later writes elide stale reads, and self-clears context between plan and execute phases. The CLI offers 3code to start, :provider add to stack another provider, -p private mode that gates turns to zero-training providers and turns the token bar magenta, a loop guard that halts runaway autonomous edits at configurable thresholds, human-readable .3log session persistence with resume, and built-in native web search. Used as a Nim library, the same agent loop, tool calls, sandbox and session persistence are exposed through return values and callbacks so you can attach a web frontend, chat bot, CI runner or IDE plugin.