Codebuff
Also known as: Codebuff AI, Freebuff, Manicode, CodebuffAI, Codebuff, Inc.
Open source terminal coding agent that splits each change across specialized agents, an orchestrator with file picker, thinker, editor and reviewer agents, with per agent model choice on OpenRouter, a TypeScript SDK for embedding, and a public Agent Store of reusable agents.
Codebuff is an open source, terminal native AI coding agent whose distinguishing idea is a multi agent architecture. It coordinates specialized agents that each own a step, not one model doing everything. An orchestrator called Buffy reads the prompt and gathers context, file picker and code searcher agents find the relevant code, a thinker works through hard problems, an editor writes the change, and a reviewer checks it.
The pitch is that splitting the work this way gives better context, more accurate edits, and fewer errors than single model tools, especially on architectural, multi file changes. The project is published under Apache-2.0 with a public repository, now named freebuff, that carries roughly 13.7k stars and more than 12,000 commits.
Codebuff, Inc. was founded by James Grugett and Brandon Chen in San Francisco and backed by Y Combinator in its Fall 2024 batch. The product launched originally as Manicode in 2024 and was renamed to Codebuff after a trademark dispute.
Codebuff is built for wholesale changes across a codebase, not single file autocomplete. A Tree-sitter scan maps the repository, edits land across many files from one prompt, and typechecks and tests run after the changes. It lives in the terminal, installs through npm, and works with any tech stack. Four modes set the depth of each run. Max runs several editors in parallel and picks the best result, Plan writes a plan without touching files, Lite is cheaper and faster, and Default sits between them.
A knowledge.md file scaffolded by /init gives the agents project context, and Codebuff also reads AGENTS.md and CLAUDE.md files and loads skills from .claude/skills folders, so a Claude Code setup carries over. Its homepage claims higher quality output and runs more than 100 seconds faster than Claude Code, citing BuffBench, its own suite of more than 175 tasks, a self reported result on its own harness.
Three strengths stand out for developers. On model choice, each agent definition takes any OpenRouter model, so different roles in one workflow can run on models from different makers, and Codebuff's own modes pair a Claude Opus orchestrator with GPT and Gemini subagents.
On extensibility, a TypeScript SDK exposes a CodebuffClient whose run method takes an agent id, a prompt, custom agent definitions, custom tool definitions and an event callback, so teams embed the agents in their own applications and CI/CD.
On reuse, a public Agent Store at codebuff.com/store lists about 160 agents that others run, compose or fork, with custom agents authored as TypeScript definitions declaring their model, their permitted tools and the subagents they may spawn.
Freebuff, Codebuff's free sibling, is ad supported, with no subscription or card and a daily usage allowance. It now ships as a CLI, a desktop app in beta, a web app builder, a cloud agent for GitHub repositories and a chat app, on DeepSeek, GPT, MiniMax, MiMo and GLM models, and its desktop app can also run a locally installed Claude Code or Codex on the user's own accounts. Its privacy terms say prompts and messages, including pasted content, may be analyzed to personalize those ads, though connected repositories and separate uploads are excluded from advertising providers.
The main caveats are that Codebuff itself is terminal only, leaving desktop, web and chat to Freebuff, that custom agents require TypeScript, that no security attestation, SSO or governance surface is documented, and that its own comparison page points buyers who need SSO, RBAC, compliance programs or private networking to Claude Code. Despite the open source license, the agent authenticates against a Codebuff issued API key and reaches hosted models, so prompts and code context leave the machine in normal operation. Codebuff has also deprecated connecting a Claude Pro or Max subscription.
From our comparison video
Claude Code vs Codebuff: 7 Questions for Engineering Teams Picking a Coding Agent
Our 8 minute video puts Codebuff side by side with Claude Code. It covers what each one does and how each is priced, then when to choose which.
Recorded in October 2026. The profile here is live and updates as research is verified.
Vendor details
Canonical URL
https://www.codebuff.com
Category
Coding agent
Subcategory
Open source multi-agent terminal coding agent
Funding status
Independent. Codebuff, Inc., founded by James Grugett and Brandon Chen, is based in San Francisco and backed by Y Combinator in the Fall 2024 batch. It launched as Manicode in 2024 and was renamed Codebuff after a trademark dispute. The Apache-2.0 repository, now named CodebuffAI/freebuff, carries roughly 13.7k stars, 1.5k forks and more than 12,000 commits as of October 2026, alongside the paid Codebuff subscriptions and the free, ad supported Freebuff apps.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Tool calling is deep within the developer's environment: a documented run_terminal_command tool, file reading tools, git analysis and commit workflows, project test execution, bash mode invoked with a leading exclamation mark or /bash, and built-in web research and browser use. MCP servers added in .agents/mcp.json over stdio, HTTP or SSE become tools for all base agents, with Notion, GitHub and remote API examples documented. The TypeScript SDK accepts customToolDefinitions so embedders add their own tools, and agents embed into applications and CI/CD pipelines through it. A public Agent Store at codebuff.com/store publishes reusable agents. Model access is brokered through OpenRouter, with model selection set per agent definition.
In practice
You need a change touching many files, endpoints, and tests at once. Codebuff's agents map the codebase, make the edits from one prompt, and run your tests to verify nothing broke.
You want to try an autonomous coding agent without paying for one. Freebuff is free and ad supported across several models, with no card or API key, and the Apache-2.0 source lets you inspect how the agents work.
You want coding assistance inside your own product or CI pipeline. Codebuff's TypeScript SDK lets you embed its agents, define custom tools, and route each task to whichever model fits the job.
Sources & related URLs
Related / legacy domains
Research sources
Agentic Index coverage score
9.0 / 14 capabilities · 64%
| Integrations & Tool Calling | Full |
|---|---|
|
MCP servers added in .agents/mcp.json over stdio, HTTP or SSE become tools for all base agents, with Notion, GitHub and remote API examples, alongside a basher agent for terminal commands, a researcher for web search and custom tool definitions through the SDK. Sourcecodebuff.com/docs/tips/mcp-servers and docs/agents/overviewread 2026-10-24 |
|
| Workflow Orchestration | Full |
|
The orchestrator, Buffy, reads the prompt, gathers context and spawns specialized agents (file picker, code searcher, researcher, thinker, editor, reviewer and basher), and agents spawn others listed in their spawnableAgents configuration. Max mode runs several editors in parallel with different strategies, and a selector picks the best result. Sourcecodebuff.com/docs/advanced/how-does-it-work, docs/agents/overview and docs/tips/modesread 2026-10-24 |
|
| Knowledge Grounding & RAG | Partial |
|
A Tree-sitter scan builds a code map of the repository, file picker and code searcher agents find the relevant code, and knowledge.md, AGENTS.md or CLAUDE.md files carry conventions and commands. Context is assembled per run, with no persistent index, graph or embeddings layer described. Sourcecodebuff.com/docs/advanced/how-does-it-work and docs/help/faqread 2026-10-24 |
|
| Human Oversight & Guardrails | Partial |
|
Each agent definition limits the tools it may use and the agents it may spawn, a customer set constraint on what the agent can do, and Plan mode gathers context, asks clarifying questions and writes a plan without touching files. A reviewer agent checks changes after they are made. There is no step where a person approves an edit or command before it runs. Sourcecodebuff.com/docs/tips/modes and docs/agents/agent-referenceread 2026-10-24 |
|
| Security, Identity & Governance | Not documented |
|
There is no certification, SSO, role based access or audit log, and Codebuff's own comparison page names SSO, RBAC, compliance programs and private networking as reasons to choose Claude Code. Its privacy summary, which it says applies to Codebuff and Freebuff, says Freebuff uses messages, agent traces and files to run the service and may analyze them to personalize ads, with training use only where a model or feature says so. Sourcecodebuff.com/docs/advanced/privacy and docs/advanced/claude-code-comparisonread 2026-10-24 |
|
| Observability & Auditability | Partial |
|
The SDK's event callback streams agent start and finish, tool call and tool result, text and error events as a run executes, so an embedder can watch each step, but keeping that record is left to the embedder and no audit log, retained trace or export is offered. Sourcecodebuff.com/docs/advanced/sdkread 2026-10-24 |
|
| Memory & State Persistence | Partial |
|
The SDK's previousRun parameter carries conversation state from one run into the next, and knowledge.md files load every run as project instructions, written by the user or generated by /init, with AGENTS.md and CLAUDE.md read in their place. There is no separate memory the agent keeps for itself with a stated scope and lifetime. Sourcecodebuff.com/docs/advanced/sdk, docs/tips/knowledge-files and docs/help/faqread 2026-10-24 |
|
| Deployment & Data Residency | Partial |
|
The CLI runs in the developer's terminal, edits and commands run locally, and the Apache-2.0 source can be built and inspected, but the agent authenticates with a Codebuff API key and reaches hosted models, so prompts and code leave the machine in normal use. There is no customer hosted deployment or region choice, and the FAQ points to a Docker container for running the agent against a scoped copy of the code. Sourcegithub.com/CodebuffAI/freebuff and codebuff.com/docs/help/faqread 2026-10-24 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
A public Agent Store at codebuff.com/store lists about 160 published agents to run, compose or fork, and eight named built in agents each do their own job (editor, reviewer, thinker, researcher, file picker, basher, code searcher and the base orchestrator), each a whole unit on its own. Skills in .agents/skills or .claude/skills folders load on demand and run as slash commands. Sourcecodebuff.com/store, docs/agents/overview and docs/tips/skills; codebuff.com/docs/agents/overviewread 2026-10-24 |
|
| Triggers & Channel Coverage | Partial |
|
Agents start from the terminal CLI, by @mention inside a session, or from the customer's own code through the SDK, which Codebuff positions for CI/CD pipelines and automation scripts, with a GitHub Actions example that runs on pull requests. Codebuff itself offers no schedule, webhook or event trigger that wakes an agent on its own. Sourcecodebuff.com/docs/advanced/sdkread 2026-10-24 |
|
| Model Flexibility & Routing | Full |
|
Each agent definition takes any OpenRouter model string, so agents in one workflow can run on models from different makers. Codebuff's built in modes already mix them, with a Claude Opus orchestrator in Default, Max and Plan, GPT-5.6 Luna in Lite, and Gemini Flash Lite subagents for file discovery, terminal commands and research. Freebuff offers DeepSeek, GPT, MiniMax, MiMo and GLM models. Sourcecodebuff.com/docs/agents/agent-reference, docs/advanced/what-models and docs/tips/modesread 2026-10-24 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
The @codebuff/sdk package exposes a CodebuffClient with an API key and working directory and a run method taking an agent, a prompt, previous run state, custom tool and agent definitions, knowledge files and an event callback, meant for CI/CD pipelines, batch jobs, VS Code extensions, web apps and automation scripts. Custom agents are authored in TypeScript and published through the Agent Store. Sourcecodebuff.com/docs/advanced/sdkread 2026-10-24 |
|
| Testing, Debugging & Optimization | Partial |
|
Codebuff publishes BuffBench, its own suite of more than 175 tasks rebuilt from open source commits, to compare itself with Claude Code, and a reviewer agent checks changes after each edit. There is no stated way for a customer to run evaluations against its own agents or workload, so the harness is not a customer surface. Sourcecodebuff.com, docs/tips/modes and github.com/CodebuffAI/freebuffread 2026-10-24 |
|
| Browser & Computer Use | Partial |
|
Codebuff's repository README says its agents investigate documentation and test apps in a browser, but nothing describes what the agent can do there, so the capability is claimed, not shown. Sourcegithub.com/CodebuffAI/freebuffread 2026-10-24 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Codebuff from $100 per month, or pay as you go at $0.01 per credit. Freebuff is free and ad supported.
Credits spent by task complexity, bought through monthly subscriptions ($100, $200 or $500) or pay as you go at $0.01 each. Freebuff is free and ad supported.
Included quota
Codebuff subscriptions give 1x usage for $100, 2.5x for $200 and 7x for $500. Freebuff is free within a daily usage allowance set out in its README, with models billed against it by a multiplier shown in the picker.
What is public
Subscription prices, their usage multiples and the $0.01 credit price are published on codebuff.com/pricing, and Freebuff's free terms on freebuff.com and in its README. Enterprise pricing is not published.
Billing mechanics
Credits are spent by task complexity. Subscriptions bundle usage at 1x, 2.5x or 7x, and pay as you go credits cost $0.01 each without a subscription. Freebuff is funded by ads.
Cost watchouts
Multi agent runs consume more credits per task, and the largest subscription reaches $500 a month.
Variable cost rationale
Codebuff meters credits, spent by task complexity, and multi agent runs use more per task, so spend tracks usage rather than a flat seat. A subscription bundles usage at 1x, 2.5x or 7x for $100, $200 or $500 a month, and anything beyond is bought pay as you go at $0.01 a credit, with no published cap. Freebuff costs nothing, and teams that build on the open source CLI pay their own model costs instead.
Additional watchouts
Multi agent runs spend more credits per task than single agent work, and Freebuff uses prompts, messages, agent traces, code and repository data to provide its service, which is funded by advertising.
Overage / add-ons
Credits beyond a subscription's usage are bought pay as you go at $0.01 per credit. Simple queries cost less and complex changes more.
Sales call required
No, self serve available
Free / trial
Free through Freebuff, ad supported, with no credit card and a daily usage allowance.
Lowest paid plan
Pay as you go at $0.01 per credit with no subscription, or subscriptions from $100 a month
Commercial notes
There are three paths. Freebuff is free and ad supported, Codebuff is paid through subscriptions or pay as you go credits, and the Apache-2.0 source serves teams that build the harness themselves.
Key ambiguities
The credit volume behind each subscription's usage multiple is not stated, so cost per task is hard to predict in advance.
Cancellation / refund
Subscriptions can be canceled anytime. Freebuff and pay as you go need no commitment.
Support SLA / resale
Community support for the open source project. Enterprise plans are available on request.
Missing data
How many credits a subscription's 1x usage includes, and enterprise pricing, are not published.
Related vendors
- Cognition — Maker of Devin, an autonomous AI software engineer
- 10Web — Agentic website platform whose AI agents build, host and manage…
- AgentUI — Managed AI app builder for operations teams that generates and hosts…
- Aider — Open source, model agnostic terminal coding agent that edits your…
- Amp — Amp is a coding agent from Amp Frontier Corporation that runs each…
- Anthropic Claude Code — Anthropic's agentic coding system across terminal, desktop, IDE, web…
Alternatives to Codebuff
The closest documented capability profiles to Codebuff among coding agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Vibecode10.0 / 14Fuller documented coverage on Testing, Debugging & Optimization and Browser & Computer Use
- Anything11.0 / 14Adds documented Security, Identity & Governance
- Blink10.0 / 14Adds documented Security, Identity & Governance
- Builder.io10.0 / 14Adds documented Security, Identity & Governance
- Choicely9.0 / 14Adds documented Security, Identity & Governance
- v0 by Vercel11.0 / 14Adds documented Security, Identity & Governance
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded