Anthropic Claude Code vs Codebuff (2026)

MCP servers configured in .agents/mcp.json over stdio, HTTP or SSE become tools for all base agents, with Notion, GitHub and remote API examples documented. Alongside them sit a bash agent for terminal commands, a researcher for web search and custom tool definitions through the SDK.

A base agent orchestrates named role agents, agents spawn others listed in their spawnableAgents configuration, and conditional branching and programmatic step control are documented, which is multi-agent execution.

The GitHub Action runs on any GitHub event or cron schedule without a mention, beside @claude mentions and terminal, desktop, IDE, web and headless runs.

Agents start from the terminal CLI, by @mention in a session, or from the customer's code through the SDK, including in CI. What holds it short of Full: no IDE extension, web console, chat channel, scheduled run or event-driven trigger is documented.

File picker and code searcher agents map the relevant parts of a project before editing, and knowledge.md files carry conventions and commands. What holds it short of Full: that context is assembled per run, with no persistent index, graph or embeddings layer documented.

Auto memory, written by Claude as it works and scoped per repository, persists across sessions until edited or deleted and is browsable and editable through /memory.

Two mechanisms carry state: the /history command browses and resumes past conversations, and the SDK's previousRun parameter carries state from one run into the next. What holds it short of Full: that is session resumption and static project context, not memory that accumulates or learns across sessions.

The per-agent tool allowlist is a real scoping control: each agent definition limits the tools it may use and the agents it may spawn. What holds it short of Full: the vendor markets fewer confirmation prompts as an advantage, and no per-action approval gate, permission prompt or runtime policy is documented.

No certification, single sign-on, role-based access or audit log is documented. The Freebuff FAQ states that prompts and messages, including pasted content, may be analyzed to personalize advertising and that submissions may be retained for AI training where a model or feature says so; connected repositories and separate uploads are excluded from advertising providers. Open source inspectability and a published SECURITY.md are transparency rather than a security posture.

OpenTelemetry exports metrics, events and traces with a decision event for every permission prompt and a result event for every tool call, and cloud sessions log all operations for audit.

The SDK's handleEvent callback gives per-action visibility as a run executes. What holds it short of Full: it is a streaming callback the embedder must persist themselves; resumable conversation history is a transcript rather than an audit record, and no audit log, retained trace or governance surface is documented.

Claude Code runs on the developer's own machine by default, cloud sessions run in isolated Anthropic VMs or on the organization's self hosted environments, and data is encrypted in transit but not at rest.

The CLI runs locally and the Apache-2.0 source can be built and inspected, but agent definitions address models through OpenRouter, the CLI authenticates with an API key issued at codebuff.com/api-keys, and no local model, offline mode or self-hosted backend is documented in the README, docs or FAQ.

Plugins bundle skills, subagents, hooks and MCP servers and are browsable in Anthropic's official marketplace, beside bundled skills and built in subagents.

A public Agent Store at codebuff.com/store lists agents to browse and adopt; the README tells users to compose published agents from it and asks the community to publish specialized agents there. Eight named built-in agents also each do their own job.

Every model is Anthropic's: customers choose a Claude model and where it is served, but not the model maker.

The model is a field on each agent definition, so different roles in one workflow can run on different models, which is routing the customer controls. Model access is brokered through OpenRouter with a vendor-issued API key, so this is broad choice rather than bring-your-own endpoint.

The @codebuff/sdk package documents a CodebuffClient whose run method takes an agent, a prompt, previous run state, custom tool definitions and an event callback, for CI/CD, batch jobs, editor extensions and web apps. Custom agents are authored in TypeScript and shared through the Agent Store.

claude plugin eval runs prompt suites in fresh isolated sessions, grades them against a no plugin baseline and can gate CI on the score.

A reviewer agent checks changes, and the repository publishes an evals directory with a suite of more than 175 tasks that the vendor uses to benchmark its own agent. What holds it short of Full: no page shows how a customer points that harness at its own agents or workload.

The Chrome integration lets the agent operate a visible browser session with the user's logins, and computer use extends this to native macOS apps.

The repository states agents can investigate documentation and test applications in a real browser. What holds it short of Full: that is a single feature line, with no page describing what the agent can do in the browser. Shell commands, git and file operations are programmatic and do not count here.