AI coding agents move to the terminal, the phone and the budget line
Two new open source projects put an AI coding agent in the terminal and another on your phone, while a third sells per-developer spending caps. Together they show where the developer tool fight has moved.

One arrival is Docent, a Rust CLI that chats with PDFs and Office files, searches the web and connects MCP servers, all from the terminal. The repository lists prebuilt binaries for Linux, macOS and Windows. The pitch is local first: if no agent is running, the CLI downloads the framework and the models it needs and starts one itself.
That last detail is the part worth watching. The binary carries a built-in agent, SLP framework 1.2.9, and the first run pulls it down in the background and logs to ~/.smartloop/server.log. The built-in agent uses GPU inference through Metal on macOS and Vulkan on Linux and Windows, which means a laptop with a decent GPU can do the work without sending a prompt to a vendor. For developers under data rules, that changes what is allowed. What is not clear from the repository is how large the downloaded models are or how good they are at code.
The project also keeps its old name. It installs as both docent and smartloop, and the repository says the two are the same binary. Existing commands keep working, which matters if anyone scripted the earlier release.
A phone for asking, a PC for merging
Nautilus splits the coding agent in two. A phone PWA handles the prompt, streams the agent's turn and shows a live preview. A desktop app then shows a real content diff and merges the work into the project. The repository states that it never touches your .git directory.
The setup runs on free infrastructure: a Lightning AI Studio runner, a public HTTPS port carrying the app, the API and the preview, and no credit card, domain, Docker or paid VM. The free OpenCode Zen models work straight away. The README is candid about the trade-off: a phone is a good place to ask for work and a bad place to review it. The PC only has to be reachable when you ask it to sync.
Neither project publishes benchmarks, user numbers or funding. Both are new projects, and the claims about model quality and cost come from their own documentation.
Kimchi sells the cap, not the model
Kimchi takes the opposite position on trust. Its site describes one enforcement layer for spending, data, deployment and model routing. Caps can be set by user, team, API key or model, and the site says requests stop at the limit, not after. The example on the page shows a user budget of $820 against an $800 hard stop.
Deployment options run from serverless to a VPC to on-premises, and the site says prompts, code and outputs are excluded from model training. It also claims an inference endpoint for open source models with latency around 80ms. Those figures are the company's own and have not been independently tested.
The reason budget control is now a product is simple. Agents run for minutes, not seconds, and a single session can burn through a team's monthly allowance before anyone notices. Per-key caps also matter for CI pipelines, bots and contractors, the places where a leaked credential turns into a bill. Kimchi does not say what it charges, and no pricing appears on the page.
Where the assistant stops and the platform starts
All three projects sit inside a larger argument about who controls the tools developers use. A post on Eval.blog lays out the case against Google's Android Developer Verification Program, which the author says turns Android into a locked-down platform on 1 January 2027. The post notes that verification applies to every app on a certified device, not only to apps in the Play Store. It also cites F-Droid warning that developer verification will end free and open source app distribution as it is known today.
Right now, APKs just work.
That is the author's line, and the argument is theirs. But the same logic applies to developer tools. A coding agent that runs on your machine, or on a runner you control, does not depend on a vendor's terms of service. A tool that enforces budgets through a hosted layer does.
The counterpoint is maintenance. Docent compiles llama.cpp on install and needs Rust, CMake and a C/C++ toolchain; the Linux build bundles the Vulkan loader, and macOS builds are signed and notarized while Windows builds are signed. That is a lot of moving parts for a tool that has to keep working after an upstream change.
The web platform argument, again
Nolan Lawson's 3 October post on why developers do not use the platform is a useful frame here. He argues that browsers spent years catching up with libraries, and that habit stuck. Developers reach for the tool they know, even when the browser already does the job.
The same reflex now applies to agents. Teams reach for the hosted assistant they already pay for, not the CLI they would have to compile. Lawson also notes that documentation shaped the choice: npm packages had better examples than the platform did. Open source agent projects face the same problem, with a README as their only shop window.
Two smaller projects show how wide the field has become. Herbarium keeps the HTML pages that AI tools generate in a plain folder of files and brings them back for spaced review. The 16agents project asks a coding agent to take a personality test, with a self-awareness score and four hidden traps to catch agents that claim behaviour they do not show. Its authors call it a horoscope with a regex. Neither is a coding tool in the usual sense, but both point at the same habit: developers now keep, sort and audit what their assistants produce.
What to watch
The practical question for a team this quarter is not which model wins. It is where the agent runs and who can stop it. Docent and Nautilus answer that with local or free infrastructure and no account. Kimchi answers it with a control plane and a hard cap.
None of the three has published adoption figures, and none has said how much it costs to run at scale. Until they do, the choice comes down to the constraint you already have: a data rule, a bill, or a laptop that has to stay closed.
Sources
8- 01Docent: an open source private AI assistant in your terminalEN
- 02Nautilus: run a coding agent from your phone, merge on your PCEN
- 03Kimchi: governed AI coding platform with hard budget capsEN
- 04Authoritarian tendencies of Android Developer VerificationEN
- 05Why don't more developers use the platform?EN
- 06Herbarium: keep and review the HTML pages AI tools generateEN
- 0716agents: a personality test for your agentEN
- 08OpenBSD developers reject uutils coreutils portEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.