Claude Code Best (CCB): a buildable, self-hostable Claude Code fork

What Claude Code Best actually is, and who it is for

The README describes CCB as a full engineering reconstruction of Anthropic's official Claude Code, with the added claim that it implements enterprise-tier features and features that otherwise require logging in with a Claude account, plus extra features of its own. The package.json describes it as a reverse-engineered Anthropic Claude Code CLI, an interactive AI coding assistant in the terminal. The npm package name is claude-code-best and it declares three binaries: ccb, ccb-bun and claude-code-best.

The intended reader is someone who already uses Claude Code in a terminal and wants to keep going when the official binary is not the right shape. The README states that CCB is fully compatible with existing Claude Code configuration, so original config files do not need editing, and that Dynamic Workflow and Goal are present. That compatibility claim is the whole pitch: same config surface, more moving parts around it.

It is not for people who want a hosted product with a support contract. The repository is public, the code is TypeScript, and the extras are the point. If you are happy with the official CLI and its account requirements, the value here is mostly the self-hostable pieces: remote control, artifacts, monitoring.

How CCB works: Bun runtime, workspace packages, and an agent loop with hooks

The runtime is Bun. package.json sets engines.bun to >=1.3.0, and every script in the file goes through bun: build, dev, test, lint. There is a build.ts at the repository root and a vite.config.ts, so there are two build paths, exposed as build:bun and build:vite. The published artifact is dist, with dist/cli-node.js and dist/cli-bun.js as the entry points behind the three bin names. A postinstall script ships in the files list, which matters if you install from npm into a controlled environment.

The repository is a Bun workspace with three globs: packages/*, packages/@ant/* and packages/@anthropic-ai/*. The @anthropic-ai scope suggests the original package layout is mirrored rather than flattened, which is consistent with the compatibility claim. Top-level directories include src/, spec/, tests/, scripts/, docs/, vendor/ and teach-me/, plus AGENTS.md, CLAUDE.md and DEV-LOG.md.

Around the loop sit the features that distinguish CCB. /goal takes an objective and drives the agent across turns until completion, with a token budget, completion and blocked audits, pause, resume, continue and clear subcommands, and automatic pausing on network interruption. The source lives in src/commands/goal/ and src/services/goal/. /ultracode injects a workflow orchestration manual and exposes a Workflow tool that runs deterministic JavaScript with agent, pipeline, parallel and phase primitives, plus a two-column /workflows monitor, journal replay, token budget and a concurrency cap. Pipe IPC lets multiple instances cooperate over a local socket on one machine, with LAN discovery between machines and a /pipes selection panel. ACP support is advertised for Zed and Cursor, including session resume, Skills and permission bridging.

Installing CCB and running a first /goal loop

The README does not print an install command, and it does not document a global install or a version flag. What package.json supplies is the identifiers: the package name is claude-code-best, the binaries are ccb, ccb-bun and claude-code-best, and the engines field requires Bun >=1.3.0. The repository also ships scripts, and the ones named in package.json are how you get a build on your own machine:

bun install
bun run build
bun run dev

Those three keys come straight from the scripts block: build runs build.ts, dev runs scripts/dev.ts, and the workspace is installed with bun install because the repository is a Bun workspace. The README does not state what output to expect from each, so treat the dev script as the interactive entry point and the build script as the packaging step.

For model access, the README lists custom model providers under /login, with OpenAI, Anthropic, Gemini and Grok named as compatible. That is the step to complete before expecting a first answer, and it is documented as a command rather than a config file edit:

/login

The first genuinely CCB-specific thing to try is a goal. The README documents the subcommands pause, resume, continue and clear, and says the loop pauses automatically when the network drops. The command form is given in the README as /goal with an objective:

/goal <objective>

Expect the agent to keep working across turns until it reports completion or blocks. If you would rather watch the orchestration than a single loop, /ultracode plus /workflows opens the two-column panel; the README points at https://ccb.agent-aura.top/docs/features/workflow-scripts for the script API.

Where CCB gets awkward: licence, cadence and the reverse-engineering premise

The licence is the first thing to resolve. The repository has a LICENSE file and the README renders a GitHub licence badge, but the badge target is all the metadata shows; the licence identifier itself is not stated in the repository description. That is not a small gap. This project is described as a reconstruction of a commercial vendor's CLI, so the terms under which it is distributed are a decision input, not a footnote. Read LICENSE before you plan anything around it.

The second constraint is the upgrade surface. Three releases landed in under a month: v2.8.2 on 2026-06-25, v2.8.3 on 2026-07-09, v2.8.4 on 2026-07-21, and the default branch saw a push on 2026-08-24. That is a fast-moving tree. Teams that pin dependencies and audit changes will spend real time on it, and the README does not document a rollback procedure or a long-term support branch.

The third is scope. The feature table is long, and each row points at a separate documentation page: artifacts with Cloudflare Worker and R2, Langfuse monitoring, Sentry, GrowthBook, Channels over MCP, voice input, computer use, Chrome automation. None of that is wrong, but it means a CCB deployment is a small platform, not a binary. If you only want an interactive assistant in a terminal, most of the table is dead weight you now have to reason about.

A last caveat that the documentation makes easy to miss: the README does not describe what happens to in-flight goal state across a version upgrade, and it does not state whether /goal budgets are persisted per project or per session.

CCB against the official Claude Code CLI

The honest comparison is with Anthropic's own Claude Code, which the README links to at docs.anthropic.com. The difference is not the core interaction. Both are terminal agents that read and write files in a repository. The difference is what surrounds the loop and who controls it.

On the official CLI, the account and the vendor decide which features you get. CCB's stated position is the inverse: it reimplements the enterprise and account-gated features and, in the README's words, closes all the external control points. Remote control is self-hosted through Docker, so the interface runs on infrastructure you own and can be opened on a phone. Artifacts are a Cloudflare Worker plus R2 that the README says is fully open source and self-hostable, with 7-day and 30-day expiry. Model access goes through /login, with OpenAI, Anthropic, Gemini and Grok listed as compatible providers, so the model is a configuration choice rather than a fixed one.

That is the real trade. You gain control of the runtime, the storage and the model endpoint. You take on the build, the upgrades, the licence question and the operational surface of every self-hosted piece. If your constraint is a vendor relationship rather than control, the official CLI is the simpler answer, and CCB's extra configuration is cost without benefit.

Maintenance, upgrades and what the licence question implies

The last push to the default branch was on 2026-08-24, and the newest release in the list is v2.8.4 from 2026-07-21. The repository is not archived. Those are the facts; anything about how the project is run day to day is not visible from here.

The practical upgrade cost sits in three places. First, Bun: engines.bun is >=1.3.0, so a CCB upgrade can require a runtime upgrade, and the runtime is not something you can pin away from indefinitely. Second, the workspace layout: packages/@ant/* and packages/@anthropic-ai/* mirror upstream names, so a change in the reconstructed surface can ripple through several workspace packages at once. Third, the extras with external dependencies. Langfuse, Sentry and GrowthBook are integrations with their own configuration pages, and the artifacts feature depends on Cloudflare Worker and R2. A CCB upgrade is therefore a small integration test, not a version bump.

On licensing, the only responsible statement is that the repository description does not name the licence. The repository carries a LICENSE file and the README shows a licence badge, so the answer is one click away in the repository, but it is not stated here. Treat that as the first thing to read, especially if you intend to distribute a build or run CCB inside a company. Nothing here is legal advice, and the reverse-engineering framing makes it worth a deliberate read rather than a skim.

Editorial conclusion

Adopt CCB if you want a Claude Code CLI you can build from source, read, and extend, and if the extras (Goal, /ultracode, Pipe IPC, self-hosted remote control) match how you already work. Do not adopt it if you need a stable, contractually supported tool, or if your team cannot absorb the upgrade cadence: v2.8.2, v2.8.3 and v2.8.4 all landed between 2026-06-25 and 2026-07-21, with a further push on 2026-08-24. Before committing, verify three things yourself: which licence the LICENSE file actually carries, whether the Bun >=1.3.0 toolchain and the build script run cleanly on your machines, and whether /login against your own OpenAI, Anthropic, Gemini or Grok endpoint behaves the way you expect.

Frequently asked questions

Official sources

1. claude-code-best/claude-code on GitHub
2. Issues
3. Project website
4. README
5. Releases