Compare
Free coding agents, compared honestly.
Most "free" terminal coding agents are not free in the same way. Some want an account, some want an API key, some want both, and one of them had its free tier switched off in June 2026. This table answers the two questions that actually decide whether you can start today — do you need an account, and do you need a key — alongside licensing, local models, memory, multi-agent routing, and what happens when you hit a limit.
OB-1 is the free terminal coding agent we build: Apache-2.0, no account, no card, no key. It is listed first because it is ours, so read the rest of the table before taking our word for anything.
| Capability | OB-1 | Claude Code | opencode | aider | Pi | Gemini CLI | Freebuff | Cline | Cursor |
|---|---|---|---|---|---|---|---|---|---|
| License | Apache-2.0 CLI | Proprietary | Open source | Open source | MIT | Gemini CLI was Apache-2.0; successor Antigravity CLI is closed source | Apache-2.0 | Apache-2.0 | Proprietary |
| Account required | No — no account and no card | Yes — a Claude account or Console credentials | No for the CLI; an account only for OpenCode Zen billing | No — but you need an account with whichever provider issues your key | No — but you need a provider account for /login or a key | Yes — a Google account, and individual free-tier access ended Jun 18, 2026 | No — no account and no card | No for bring-your-own-key; yes for ClinePass or Cline credits | Yes — a Cursor account |
| API key required | No — keyless providers answer the first message; your own keys are optional | No if you sign in on a Claude plan; yes on Console/API billing | Yes — bring your own key, or top up an OpenCode Zen balance | Yes — bring a provider key or point it at a local endpoint | Yes — an API key, or /login with a supported provider | Yes now — API-key auth or an enterprise Gemini Code Assist licence | No — no key and no card | Free options via Cline's own provider; otherwise bring a key | No — models come from Cursor's managed pool |
| Works with no API key | Yes, built-in free cloud models | Sign-in required | Free models advertised | Usually bring a key or local endpoint | No — /login or a key is required first | Was free with a Google account; free tier shut off Jun 2026 | Yes — no key, no card, funded by text ads | Free options via Cline's own provider; otherwise bring a key | Free Hobby plan with limited Agent requests |
| Bring your own endpoint | Yes, any OpenAI-compatible URL | Limited | Yes | Yes | Yes — a unified multi-provider API covering OpenAI, Anthropic, Google and others | No — Gemini/Google models only | Limited — Claude Code / Codex passthrough | Yes, any OpenAI-compatible API | Not a documented plan feature |
| Local models | Ollama, LM Studio, llama.cpp, vLLM, custom | Not supported first-party | Yes | Yes | Through its OpenAI-compatible provider support | Not supported | Not documented | Ollama, LM Studio | Not a documented plan feature |
| Persistent memory | Inspectable project memory graph | Claude-side/project memory | Project workflow features | Repo/session context | Session state and context management, not a memory graph | GEMINI.md context files | Project-aware multi-agent context | Editor session/project context | IDE project context and rules |
| Multi-agent routing | Solo, Fusion best-of-N, verified escalation, subagents | Subagents through Claude ecosystem | Multi-session and built-in agents | Architect/editor workflows | Single agent by design — subagents and plan mode intentionally omitted | Single agent with tools and MCP | Coordinated specialist agents | Single agent with MCP tooling | Agent over a managed frontier-model pool |
| Sandboxing | Seatbelt/bubblewrap modes plus permission controls | Local permission prompts + OS sandboxing | Local controls | Local controls | None built in — the docs recommend containerizing Pi yourself | Docker/Podman or macOS seatbelt sandbox | Not documented | Explicit approval for every action | IDE-level approvals and controls |
| Hosted tier | Optional paid tier — hard-capped prepaid credits, no overage path | Anthropic subscription or API billing | OpenCode Zen: a $20 pay-as-you-go balance, or bring your own keys | Bring your own keys | None — you pay your chosen providers directly | Google AI/Vertex billing beyond the now-ended free tier | None published — free and ad-supported | ClinePass $9.99/mo, or usage-based credits via Cline | Pro $20, Pro+ $60, Ultra $200 per month |
| Free tier durability | Free users get new free models after 30 days; hosted plans get them immediately | Depends on an Anthropic subscription or API credits | Free models depend on the provider connections you supply | No hosted tier — depends on the keys or endpoints you bring | No free tier — durability is whatever your provider offers | Shut off Jun 18, 2026; replaced by closed-source Antigravity CLI | Free today, funded by sponsor demand | Depends on the provider or plan you pick | Hobby stays free; paid usage bills on demand above the allowance |
OB-1 vs Claude Code
Claude Code is the polished Anthropic-native agent. OB-1 is the open, provider-neutral path for teams that want free onboarding, inspectable memory, and an exit plan from any one model vendor.
Read comparisonOB-1 vs Gemini CLI
Gemini CLI launched in June 2025 with a generous free tier — 1,000 requests a day at no charge. In March 2026 the free tier was cut to Flash-only models, and on June 18, 2026 Google shut it off entirely, pointing users to the closed-source Antigravity CLI. OB-1's free path is a router built into the CLI; new free models reach free users after 30 days and paid users immediately.
Read comparisonOB-1 vs Freebuff
Freebuff is a genuinely free, ad-supported coding agent from the Codebuff team (YC): no API key, no card, and a curated catalog of strong open models paid for by text ads between agent turns. OB-1's free path is ad-free and user-pointable — the same CLI routes to any OpenAI-compatible endpoint, including models running on your own machine.
Read comparisonOB-1 vs opencode
opencode is a strong open coding agent with a broad product surface, and it also ships free models with no key. OB-1 beats it on two things: persistent project memory and a self-checking test-passing loop, both in the free tier.
Read comparisonOB-1 vs aider
aider is a proven terminal pair programmer with excellent repo-map and git workflows. OB-1 is for longer-running projects that need memory, multi-agent routing, and a free no-key first-run path.
Read comparisonOB-1 vs Cline
Cline is a mature Apache-2.0 agent that lives where you already work: VS Code and JetBrains, plus a CLI and an SDK, with first-class MCP support and a real enterprise tier. OB-1 is terminal-first, built around inspectable project memory and hosted credits that are capped in advance rather than metered as you go.
Read comparisonOB-1 vs Pi
Pi (pi.dev, earendil-works/pi) is an MIT-licensed terminal coding agent built on a unified multi-provider LLM API, and it is deliberately narrow: no subagents, no plan mode, no built-in permission system. It also is not a free path — you run /login or supply an API key before it answers. OB-1 starts with no key at all and adds parallel subagents, Fusion best-of-N, and a self-correction loop.
Read comparisonOB-1 vs Cursor
Cursor is a complete AI-first IDE — a VS Code fork with inline tab completion, a large frontier-model pool, and Bugbot code review. OB-1 is a terminal agent, not an IDE. The sharpest practical difference is how spending ends: Cursor's paid plans include an allowance and then bill on-demand usage in arrears, while OB-1's hosted plans are prepaid and stop at the cap.
Read comparison