The strongest $0 starting stack is GitHub Copilot Free or Cursor Hobby for editor assistance, Codex on ChatGPT Free for small agent tasks, Antigravity Individual for a free Google-side terminal path, and an open-source harness such as OpenCode or Aider when you can use a local model or accept separate token charges. Continuum is genuinely free as an app: Chat, Code, worktrees, multi-device clients, and provider orchestration have no Continuum fee when you bring your own keys and subscriptions. No hosted frontier coding agent offers credible unlimited use at $0. Unlimited usually means local inference on hardware you own, unlimited completion on a paid plan, or marketing that omits a fair-use cap.
- Copilot Free is the simplest zero-dollar route to completion in an editor you already use.
- Cursor Hobby is a real non-expiring tier, but Cursor publishes no fixed usage number.
- Codex on ChatGPT Free covers quick tasks; it is not an unlimited professional allowance.
- Antigravity Individual is the current free Google terminal route; consumer Gemini CLI access ended in June 2026.
- OpenCode and Aider are free software, not automatically free inference.
- Continuum charges $0 for the app with your own providers; optional hosted inference is separate.
The five meanings of free
A useful free-tool comparison starts by naming the resource. Software, model inference, cloud compute, hardware, and review time are separate meters. Vendors can make one free while charging for another.
| Label you see | What is free | What may still cost money |
|---|---|---|
| Free completion tier | A monthly amount of inline suggestions | Chat, agents, premium models, or work after the allowance |
| Free agent tier | A small number of tool-using tasks or requests | More tasks, stronger models, cloud environments, overage |
| Open-source coding agent | The client or harness license | Model API, hardware, hosted service, operations |
| Free BYOK workbench | The application and orchestration layer | The provider key or subscription you bring |
| Local model | No vendor token invoice for inference | GPU or unified-memory hardware, electricity, setup, latency |
The practical test is to follow one prompt all the way to the model. Which software receives it? Where does the model run? Which account supplies the quota? Where does code execute? Who pays for the machine? If the answer changes halfway through the route, the pricing page's first number is not the total.
The best free AI coding tools at a glance
Terms checked against first-party pricing and product pages in August 2026. Quotas change; verify before depending on them for a deadline.
| Tool | What is $0 | Limit shape | Best use |
|---|---|---|---|
| GitHub Copilot Free | 2,000 completions and 50 chat requests per month | Monthly allowance | Autocomplete and light chat in an existing editor |
| Cursor Hobby | Agent, Chat, and Tab on Auto with no card or expiry | Limited, with no fixed public number | Evaluating an AI-first editor |
| Codex on ChatGPT Free | Quick coding tasks through the included free allowance | Plan quota | Trying a real terminal or cloud agent path |
| Antigravity Individual | Google-side coding agent with basic weekly limits | Weekly rate limits | A free terminal agent after the Gemini CLI transition |
| OpenCode | MIT agent software | Model and compute separate | Many providers or local models |
| Aider | Apache 2.0 terminal software | Model and compute separate | Small open harness with automatic git commits |
| OpenHands core | MIT core software and self-hostable paths | Model, Docker host, and operations separate | Autonomous jobs in a containerized platform |
| Continuum | Chat, Code, worktrees, clients, and BYOK orchestration | Provider limits still apply; hosted inference optional | Several agents and subscriptions in one free app |
The first four rows include some model access. The next three make the software free and leave inference to you. The last row makes the workbench free and drives whichever provider account you connect. Those are all legitimate free offers, but they solve different layers.
Do not compare the number of Copilot completions with Cursor agent requests or Antigravity weekly work. A completion predicts a short edit. An agent request can read files, call tools, and accumulate many model turns. Request counts are not a common unit.
Copilot Free and Cursor Hobby: free inside the editor
GitHub Copilot Free is the clearest starting point for someone who wants code completion without changing editors. The public allowance is 2,000 completions and 50 chat requests each month. That is enough to learn whether inline prediction helps, and too small for heavy agent work. Paid Copilot plans change the economics by making ordinary completions abundant and assigning a separate allowance to premium AI work.
Cursor Hobby is the better experiment when the editor itself is under evaluation. It requires no card, has no expiry, and includes limited Agent, Chat, and Tab access on Auto. Cursor does not publish a stable fixed request count for Hobby. Articles quoting older fast and slow request numbers describe retired pricing systems. The usage view on your account is the operative number.
| Question | Copilot Free | Cursor Hobby |
|---|---|---|
| Do I change editors? | No | Yes, to Cursor |
| Published allowance? | 2,000 completions and 50 chat requests monthly | No fixed public number |
| Does the tier expire? | No standing trial clock | No |
| Best thing to test | Completion in your current editor | Cursor Tab and the integrated editor loop |
| Poor test | Large autonomous ticket | Trying to budget professional agent volume from an unpublished limit |
Use the free tier until one of two things happens: a limit interrupts paid work twice, or you ration the feature so aggressively that you are no longer testing it. If neither happens over a month, do not upgrade. Headroom you never use has no productivity value.
Codex on ChatGPT Free: a real agent, a small rail
OpenAI includes Codex access at the free end of the ChatGPT plan ladder for quick tasks, with larger allowances on paid plans and an API-key route for metered work. That makes Codex the shortest path from “AI code generator” to an agent that can inspect a repository, edit files, run commands, and return a diff without buying software first.
The free allowance is a trial of the workflow, not a production promise. Give it one bounded issue with a test, not an open-ended request to improve the codebase. A scoped task teaches you whether you like the planning, sandbox, approvals, diff, and review loop before the quota becomes the only thing you notice.
Start from a clean repository
Commit or stash your own work. The first free task should have a reversible baseline and no ambiguity about which edits belong to the agent.
Choose a task with a binary finish
A failing test, a narrow bug, or a documented lint failure is better than a greenfield feature. State the command that proves completion.
Keep the sandbox tight
Use read-only for investigation and workspace-scoped writes for implementation. Do not bypass approvals to make a free trial feel faster.
Read the whole diff
The question is not whether Codex produced code. It is whether the accepted change saved more time than review and correction consumed.
Upgrade only from evidence
If the workflow fits and the allowance interrupts real tasks, compare the next ChatGPT plan with another agent subscription you may already hold.
The same logic applies to any plan-included agent. Existing subscriptions are sunk only until renewal, but they are the cheapest place to start. A second $20 seat should solve a failure you observed, not a comparison article's favorite model.
Antigravity replaced free Gemini CLI access
Older free-AI lists recommend Gemini CLI with a personal Google allowance. That route changed. On 18 June 2026, Gemini CLI stopped serving individual free, Google AI Pro, and Google AI Ultra accounts. Enterprise Code Assist licenses, Google Cloud, and paid API keys continue. Consumer users moved to Antigravity CLI.
Antigravity Individual has a free tier with basic weekly rate limits. Paid Google AI plans raise ceilings and can add a credit pool. The precise amount available to an account is plan and product state, so check Google's current pricing page rather than copying the former Gemini CLI figure.
Antigravity is useful in a zero-dollar stack when you want a second terminal agent or already use Google's ecosystem. It is not a reason to depend on one free rail. Keep project instructions portable and repository checks vendor-neutral so the next entitlement change is a login change rather than a workflow rewrite.
OpenCode, Aider, and OpenHands: free software routes
OpenCode, Aider, and the OpenHands core can all be installed without a software seat. The next decision is the model. A hosted frontier model creates a token bill. A local model can remove the vendor inference bill after hardware is available. A hosted free model offer can disappear or change. The harness remains, but the economics follow the route you configure.
| Tool | What stays free | Cheapest inference route | Hidden work |
|---|---|---|---|
| OpenCode | MIT client and agent surfaces | Local model or a provider free allowance | Provider setup, permissions, updates, model compatibility |
| Aider | Apache 2.0 terminal harness | Local OpenAI-compatible model server | Repository map quality, model tuning, release cadence |
| OpenHands | MIT core and local/self-hosted paths | Local model on infrastructure you control | Docker host, images, mounts, network policy, logs |
A local model is the only durable route to uncapped inference without a vendor meter, and even then “unlimited” means bounded by your machine. Context size, memory, throughput, queueing, and power impose physical limits. A smaller model may also require more retries and more review. Zero dollars per token is not zero cost per accepted change.
Pick OpenCode for the broadest daily surface, Aider for the smallest terminal harness and automatic commits, and OpenHands when autonomous containerized work is the actual requirement. Installing a platform to generate one function is free in the same way maintaining a server for a text editor is free.
Continuum is a free app, not free model inference
Continuum's pricing page separates the application from inference. The app is free with your own provider keys and subscriptions. Chat, Code, worktrees, the multi-agent workbench, and clients across supported platforms do not carry a Continuum app fee. Provider sessions still consume the plan or API account you selected.
Optional Continuum-hosted inference is a different line. The public ladder starts with Plus at $25 per month and rises through larger weekly hosted allowances. Selecting hosted inference creates a Continuum bill. Running Claude Code under your Anthropic login, Codex under ChatGPT, or OpenCode with your provider key does not route those model charges through Continuum or add a markup.
| Continuum layer | Price | What supplies the model |
|---|---|---|
| BYOK Chat and Code app | $0 | Your provider API key or supported subscription |
| Worktrees and agent workbench | $0 | The agent and account selected for each session |
| Mac, Windows, Linux, web, iPhone, and Watch clients | $0 app fee | Paired hosts and connected providers |
| Optional hosted inference | Paid plan from $25/mo | Continuum-hosted allowance |
This is useful if you already pay for more than one coding agent. The free layer consolidates sessions, worktrees, diffs, quota gauges, and local spend without selling the same model access again. It does not make a Claude, OpenAI, Cursor, or Google plan free. The claim is narrower and verifiable: no Continuum charge for the app and BYOK workbench.
There is no honest unlimited hosted frontier agent for $0
Search demand for “best AI for coding free unlimited” is understandable. The phrase describes a product whose marginal cost would be paid by somebody else forever. Frontier model inference, cloud sandboxes, storage, and network all have variable cost, so durable hosted products impose a request quota, usage window, credit pool, model downgrade, fair-use policy, wait queue, or paid overage.
| “Unlimited” claim | What it usually means |
|---|---|
| Unlimited completions | Short predictive edits on a paid plan; agent work uses another pool |
| Unlimited free chat | Fair-use controls, model routing, or capacity-based throttling still apply |
| Unlimited local coding | Your hardware is the cap and you paid for it |
| Unlimited open-source agent | The software loop is uncapped; the selected model may not be |
| Unlimited trial | A temporary acquisition offer that can change without fitting production |
The correct goal is predictable access, not the word unlimited. A free monthly allowance can be predictable. A paid hard-cap subscription can be predictable. A local model on known hardware can be predictable. A vague fair-use promise is harder to plan around than a smaller published quota.
Free code generation is not the same as free repository work
A free chatbot can generate a function, regex, SQL query, test fixture, or small component. That satisfies many searches for a free AI code generator. It does not mean the tool can locate the right file, preserve an interface, run the repository's checks, observe a failure, and revise the change.
| Task | Free generator is enough? | Agent adds value when |
|---|---|---|
| Draft a standalone function | Usually | The function must fit existing types, tests, and conventions |
| Explain an error message | Usually | Diagnosis requires reading several project files or running commands |
| Generate a small UI component | Often | The component must match a design system and be verified in the app |
| Fix a repository bug | Rarely | The agent can reproduce, edit, test, and report |
| Migrate an API across packages | No | The loop must search callers, update tests, and catch regressions |
Use free chat for isolated generation and reasoning. Use a free editor allowance when you remain in control of file placement. Use a free agent allowance when the tool loop is the thing being evaluated. Paying for an agent to produce a snippet is wasteful; pasting a whole private repository into a free chat to avoid an agent plan can be worse.
Privacy and security do not become optional at $0
A zero-dollar price can be subsidized by product acquisition, a paid upgrade path, model improvement terms, an existing platform business, or local hardware. Read the data and retention terms for the exact account tier. Enterprise promises do not automatically apply to a personal free account.
- Do not paste production secrets. Free or paid, the model does not need raw credentials to explain code.
- Keep repository access narrow. A GitHub app installed for one test repository should not receive every organization repository by default.
- Use a worktree or clean branch. Free allowance is not a reason to let trial edits mix with uncommitted work.
- Leave deploy and merge outside the agent. A free trial needs less authority than the normal engineering identity, not more.
- Check training and retention terms. Account type, product surface, and organization settings can change the answer.
- Remove the trial cleanly. Revoke tokens, uninstall extensions, remove MCP servers, and inspect configuration files after the test.
Open-source and local-model routes can keep prompts away from a hosted model, but only when the complete route is local. A local client pointed at a cloud API still sends the context to that provider. A self-hosted agent with a web-search tool still sends queries outside. Draw the data path instead of relying on product adjectives.
Build a $0 stack by job
| Goal | $0 stack | First reason to pay |
|---|---|---|
| Learn AI completion | Copilot Free in the editor you already use | You exhaust completion or chat allowance during real work |
| Evaluate an AI-first editor | Cursor Hobby | Agent limits interrupt tasks and Cursor Tab has already earned the migration |
| Try terminal delegation | Codex on ChatGPT Free, then Antigravity Individual as a second route | A plan cap interrupts accepted work repeatedly |
| Own the harness | OpenCode or Aider plus a local model | Frontier-model quality saves more review time than the API costs |
| Try autonomous self-hosting | OpenHands core on an existing Docker host plus a local model | Operations or model quality becomes the bottleneck |
| Run several existing agents | Continuum with your current provider accounts | Only if you choose optional hosted inference |
The cheapest stack is not necessarily the one with the most free logos. It is the one that reuses an editor, account, machine, and workflow you already have, while keeping the test reversible. Start with one tool per job. Two completion extensions in the same editor and three free agents on the same repository create noise before they create evidence.
After two weeks, count accepted tasks, review minutes, interruptions, and any provider cost. Pay when a specific free constraint costs more than the next tier: lost flow, weak model quality, missing policy, or operations you no longer want to own. Keep the free tier when it remains sufficient. Software does not deserve an upgrade for being used responsibly.
Questions people ask
Copilot Free for completion in an existing editor, Cursor Hobby for evaluating an AI-first editor, Codex on ChatGPT Free for quick agent tasks, and Antigravity Individual for a free Google-side terminal route. OpenCode or Aider is the open-source route when you supply the model.
For standalone snippets, a free chat assistant or Copilot Free is enough. For repository work, use a tool-using allowance such as Codex on ChatGPT Free or Antigravity Individual so the agent can read files, run checks, and revise.
Not as a durable hosted frontier agent. Free services impose quotas, windows, credits, routing, or fair-use controls. A local model can run without a vendor request cap, but your hardware, memory, throughput, power, and model quality are the limits.
Yes. Copilot Free includes 2,000 code completions and 50 chat requests per month as of August 2026. Paid plans expand premium AI use and make ordinary completions a much larger part of the product.
Yes. Cursor Hobby has no card requirement or expiry and includes limited Agent, Chat, and Tab use on Auto. Cursor does not publish a stable fixed request count, so check the usage view on your account.
Not for individual consumer accounts. Gemini CLI stopped serving free, Google AI Pro, and Ultra users on 18 June 2026. Consumer users moved to Antigravity CLI, whose Individual tier has basic weekly limits.
Their software is free and open source. Inference is separate. A local model can avoid a provider token bill; a hosted model uses an API key or paid route. Hardware and maintenance still count.
The Continuum app is free with your own provider keys and subscriptions, including Chat, Code, worktrees, and supported clients. Optional Continuum-hosted inference starts on a separate paid plan. Provider sessions still consume the provider account you chose.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.