On 28 August 2026 the most starred dedicated coding agent on GitHub carried ~202k stars, and two projects still recommended in most "best open source" lists were archived and read-only. A third, Continue, states in its own README that the repository "is no longer actively maintained", and its organisation now hosts a repo described as a "Static archive of continue.dev (acquired by Cursor)". That is the real risk profile of this category: a roundup written from memory is wrong within a quarter.
So the CodeCourier team did the boring part. For every project below we opened the LICENSE file and read the exact text rather than trusting the badge, pulled the last commit on the default branch, recorded the star count, and checked whether it needs your own API key. Everything was verified on 28 August 2026, and where a project is slowing down we say so with the date attached. This is a build-versus-buy reference for a platform owner, not a tour of the open source AI coding assistant market. For the whole field ranked on behaviour rather than deployment model, see our agentic coding tools roundup and 15 Best AI Coding Agents in 2026.
Bring us your security review
What is CodeCourier? An autonomous AI software engineering platform that turns a tracked issue into a tested pull request, with every run isolated in its own disposable sandbox, agent personas that encode how your team actually writes code, a learning engine that gets sharper on your repo over time, and engineering analytics your leads can audit, across GitHub, Jira, Linear, Slack and Sentry. Talk to us about a team rollout.
How we chose the best open source coding agent
There is no single best open source coding agent, because the deciding axis is your own operating capacity, which makes picking a coding agent open source teams can standardise on harder than the feature tables suggest. Seven criteria, all of them self-hosting questions:
- Licence, read from the file: we opened the actual LICENSE in each repository and recorded the exact text, including anything appended after the standard body. GitHub's detector reports
NOASSERTIONon any file with added lines, so the badge alone tells you nothing. Permissive and unconditional beats permissive with a revenue trigger. - Maintenance signal, with a date: last commit on the default branch, star count and archive status, all recorded 28 August 2026. A three-month-old commit is not disqualifying, but it belongs in your risk register.
- Governance: who can relicense this. A single-vendor repository can change terms in one commit; a foundation-governed project cannot do it quietly.
- Isolation you get versus isolation you build: whether the project ships a real boundary for model-generated commands. Four of the ten do. For the rest, that boundary is your headcount.
- Model portability: whether you can point it at any provider, at a gateway you control, or at open weights on your own hardware, and what switching costs. This is why most teams self-host at all.
- Operational surface: a single Rust binary installed per developer is a different on-call proposition from a control plane with agent backends, a database and a queue.
- Audit output: run logs, permission policies and approval records are the difference between a tool you can defend in review and one you cannot.
We did not rank on benchmark scores. Position moves with the model you attach, not the agent: see What Is SWE-bench?
Top 10 open source AI coding agents
The CodeCourier team builds and operates agent infrastructure daily, so we weight what you weight when the pager goes off: what breaks, who owns it, and what the licence says when legal asks.
-
OpenHands
-
opencode
-
goose
-
Cline
-
OpenAI Codex CLI
-
Gemini CLI
-
Kilo Code
-
mini-SWE-agent
-
Qwen Code
-
Aider
Now let's review each agent in detail.
1. OpenHands
OpenHands, formerly OpenDevin, is the closest thing here to a self-hosted control plane rather than a single-user tool. Its current form, Agent Canvas, is described in the repo as "the self-hosted developer control center for coding agents and automations", and it carried ~85k stars on 28 August 2026.
The LICENSE file is a clean MIT licence, copyright "2025 OpenHands contributors", with no appended conditions. Its default branch was last committed 27 August 2026. Agents run against pluggable backends: locally by default, or in Docker containers, on VMs, or inside your own company infrastructure. It runs the open source OpenHands agent out of the box and drives third-party agents over the Agent Client Protocol, so standardising on OpenHands is not standardising on one model vendor.
Isolation, orchestration and multi-agent routing you would otherwise write yourself already exist under MIT. The cost is that you now operate a control plane, not a CLI.
Highlights:
- Licence: MIT, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: high, long-running autonomous tasks and scheduled automations
- Sandbox included: yes, Docker, VM and company-infrastructure backends
- Bring your own model: yes, plus third-party agents over ACP
- Maintenance signal: ~85k stars, last default-branch commit 27 August 2026
2. opencode
opencode is the most starred dedicated open source coding agent on GitHub at ~202k stars, developed under the Anomaly organisation. It ships terminal-first with a desktop app in beta, distributed via npm, Homebrew, Scoop, Chocolatey, pacman, Nix and mise.
The LICENSE file is a standard MIT licence, copyright "2025 opencode", with nothing appended, and the default branch was last committed 28 August 2026. Its edge is provider breadth: the documented directory spans Amazon Bedrock, Azure OpenAI, Google Vertex AI, GitHub Copilot, OpenRouter, Cloudflare AI Gateway, SAP AI Core and Snowflake Cortex alongside local runtimes such as Ollama, LM Studio and llama.cpp, plus a custom-provider path. Its docs carry dedicated Enterprise, Permissions and Policies pages.
That makes it the strongest open source cli coding agent for a team that already runs a model gateway. It does not provision isolation, though: it executes in the shell you gave it.
Highlights:
- Licence: MIT, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: high, terminal agent with switchable built-in agent modes
- Sandbox included: no, permission and policy configuration only
- Bring your own model: yes, large documented provider directory including local runtimes
- Maintenance signal: ~202k stars, last default-branch commit 28 August 2026
3. goose
goose is the only project ranked here with foundation governance, and on a multi-year commitment that is the most consequential fact on this page. Its README states goose is part of the Agentic AI Foundation at the Linux Foundation, and the repository now sits under the aaif-goose organisation. It began at Block, is written in Rust, and carried ~54k stars.
The LICENSE file is the full, unmodified Apache License 2.0, and the default branch was last committed 27 August 2026, with recent work on permission-policy enforcement for tool calls. It ships a native desktop app, a CLI and an embeddable API, connects to 15 or more providers including Ollama, and can reuse existing Claude, ChatGPT or Gemini subscriptions through ACP instead of separate API keys. It publishes a GOVERNANCE.md and a custom-distribution guide for building your own branded build.
Two caveats: goose is general-purpose rather than code-specialised, so some of its surface is not what you are buying, and it runs natively on the developer's machine, so isolation is yours to write.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE file on 28 August 2026
- Language: Rust
- Autonomy: high, general-purpose agent across code, workflows and automation
- Sandbox included: no, runs natively on the host machine
- Bring your own model: yes, 15+ providers including local Ollama, plus ACP subscriptions
- Maintenance signal: ~54k stars, last default-branch commit 27 August 2026
4. Cline
Cline is the most complete open-source AI coding assistant if your rollout has to meet developers where they already work. One codebase ships a VS Code extension, a JetBrains plugin, a CLI that runs interactive or fully headless for CI, a web Kanban for running many agents in parallel, and a published @cline/sdk. It carried ~67k stars and is developed by Cline Bot Inc.
The LICENSE file is the full Apache License 2.0, copyright "2026 Cline Bot Inc.", and the default branch was last committed 28 August 2026. The Kanban surface matters most for a platform owner: each card gets its own git worktree, auto-commit and dependency chaining, a lightweight answer to parallel runs without full VM isolation. The IDE extension keeps a human approval step before file writes and command execution.
The honest limit: worktrees and approval prompts are workflow controls, not a security boundary. Governance is single-vendor, so today's Apache grant is permanent but tomorrow's commits are the vendor's to license.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: high, headless CLI mode plus parallel task board
- Sandbox included: no, approval gating and per-task git worktrees
- Bring your own model: yes, API key or Cline account
- Maintenance signal: ~67k stars, last default-branch commit 28 August 2026
5. OpenAI Codex CLI
Codex CLI is OpenAI's terminal agent, open sourced under Apache 2.0, and the second most starred project in this comparison at ~119k stars. It is written in Rust, so it installs as a single fast binary rather than a Node runtime plus dependency tree.
The LICENSE file is the full Apache License 2.0 with no additional terms, and the default branch was last committed 28 August 2026 at the highest commit volume of any project here. It ships a sandbox and approvals layer plus a separate execution-policy system, both first-class subjects in the repo docs. Authentication is either "Sign in with ChatGPT" using a Plus, Pro, Business, Edu or Enterprise plan, or an API key with additional setup.
That line is the trade. The code is Apache 2.0 and forkable; the default path to running it is an OpenAI account, and the sandbox and policy machinery assumes it. Genuinely open licence, not a route to model independence.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE file on 28 August 2026
- Language: Rust
- Autonomy: high, interactive and non-interactive execution modes
- Sandbox included: yes, documented sandbox, approvals and execution-policy layers
- Bring your own model: limited, ChatGPT plan sign-in or OpenAI API key
- Maintenance signal: ~119k stars, last default-branch commit 28 August 2026
6. Gemini CLI
Gemini CLI is Google's open source terminal agent and the answer most people are looking for when they search for a free ai coding agent. It carried ~107k stars and is written in TypeScript.
The LICENSE file is the full Apache License 2.0, and the default branch was last committed 26 August 2026, on a fix preventing SSRF in MCP OAuth metadata discovery, which tells you something useful about how the project handles security reports. The README documents two auth paths: a personal Google account with a free tier of 60 requests per minute and 1,000 requests per day and no API key management, or a Gemini API key documented at 1,000 requests per day across Gemini 3 Flash and Pro. It links a dedicated Sandboxing and Security page.
This is the cheapest possible pilot: unconditional licence, a quantified free tier, documented sandboxing. As with Codex CLI the model path runs to one vendor, and a per-day request cap is a hard ceiling on unattended batch work.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: high, interactive and scriptable terminal agent
- Sandbox included: yes, documented sandboxing and security configuration
- Bring your own model: limited, Google account or Gemini API key
- Maintenance signal: ~107k stars, last default-branch commit 26 August 2026
7. Kilo Code
Kilo Code is the pragmatic pick for a team that wants an open source ai agent for coding without first building a model-procurement process. It runs in VS Code, JetBrains IDEs and a CLI, and carried ~27k stars.
The LICENSE file is MIT and carries two copyright lines, "Copyright (c) 2026 Kilo Code" and "Copyright (c) 2025 opencode", which is worth knowing if your legal team audits provenance. The default branch was last committed 28 August 2026 on a JetBrains release candidate. Its commercial model is genuinely useful: the README states you pick from 500+ models, switch mid-task, and pay the provider's rate with zero markup, with no API keys required to start.
That removes the slowest step in a two-week evaluation. The trade: the zero-markup convenience runs through a Kilo account, so procurement will treat Kilo as a processor in the path even though the client is MIT. As an editor-resident agent it brings no execution isolation.
Highlights:
- Licence: MIT, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: medium to high, in-editor agent with a CLI mode
- Sandbox included: no, editor and CLI extension model
- Bring your own model: yes, 500+ models at provider pricing, or your own keys
- Maintenance signal: ~27k stars, last default-branch commit 28 August 2026
8. mini-SWE-agent
mini-SWE-agent is the research-grade option, built by the Princeton and Stanford team behind SWE-bench, and it is the project the SWE-agent maintainers now recommend over their own original: the SWE-agent README states it "has superseded SWE-agent" and that the general recommendation is to use it going forward. It carried 6,817 stars, by far the smallest count here, and that should not count against it.
The licence file is LICENSE.md, containing the MIT licence, copyright "2025 Kilian A. Lieret and Carlos E. Jimenez". The default branch was last committed 17 August 2026. The agent class is roughly 100 lines of Python and the README reports a score above 74% on SWE-bench Verified. What matters for a platform team is the environment abstraction: local, Docker and Podman, Singularity and Apptainer, Bubblewrap and contree, because swapping subprocess.run for docker exec is a one-line change in a codebase this small.
For batch evaluation, parallel issue attempts or an internal harness you intend to own, this is the cleanest foundation here. It is not a product and has no IDE surface.
Highlights:
- Licence: MIT, verified in the repository LICENSE.md file on 28 August 2026
- Language: Python
- Autonomy: high, designed for unattended batch issue-solving
- Sandbox included: yes, Docker, Podman, Apptainer, Bubblewrap and local environments
- Bring your own model: yes, model agnostic by design
- Maintenance signal: 6,817 stars, last default-branch commit 17 August 2026
9. Qwen Code
Qwen Code is the strongest option here for a genuinely local ai coding agent, because it treats provider protocol as configuration rather than architecture. It carried ~27k stars, is written in TypeScript, and is published by the Qwen team at Alibaba.
The LICENSE file is the full Apache License 2.0, and the default branch was last committed 28 August 2026 on terminal-UI migration work. The README documents multi-protocol support for OpenAI, Anthropic, Gemini and Qwen APIs, any third-party provider, and local models through Ollama or vLLM, switchable at runtime via /auth. Its feature table lists Auto Mode, Sandbox and git worktrees.
If you want a local AI coding assistant that never calls out of your network, this is the shortest path: an Apache 2.0 client, a local vLLM endpoint, and open weights you also control. The counterweight: runtime provider switching is only as safe as your egress policy, so pair it with network controls.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE file on 28 August 2026
- Language: TypeScript
- Autonomy: high, auto mode with git worktree support
- Sandbox included: yes, sandbox listed in the project's own feature table
- Bring your own model: yes, OpenAI, Anthropic, Gemini and Qwen protocols plus Ollama and vLLM
- Maintenance signal: ~27k stars, last default-branch commit 28 August 2026
10. Aider
Aider is the terminal pair programmer that defined the category, and at ~49k stars it still has the third largest following in this comparison. Its core ideas, repository mapping for large-codebase context and automatic git commits with sensible messages, were copied by nearly everything above it.
The licence file is LICENSE.txt, containing the full Apache License 2.0. The maintenance fact you need: the last commit on the default branch was 22 May 2026, three months before we checked, and the README's model recommendations still name Claude 3.7 Sonnet, DeepSeek R1 and Chat V3, and OpenAI o1, o3-mini and GPT-4o. It connects to almost any LLM including local models, covers 100+ languages, and folds linting and testing into its edit loop.
Aider is not archived and the Apache 2.0 grant on existing code is irrevocable, so nothing stops you running it. But if this is a two-year dependency, put that commit date in the risk register next to the licence and decide whether you would maintain a fork.
Highlights:
- Licence: Apache License 2.0, verified in the repository LICENSE.txt file on 28 August 2026
- Language: Python
- Autonomy: medium, interactive terminal pair programming with a watch mode
- Sandbox included: no, operates directly on your working tree
- Bring your own model: yes, almost any provider including local models
- Maintenance signal: ~49k stars, last default-branch commit 22 May 2026
Projects we checked and did not rank
A 2026 roundup listing a dead repository is worse than useless. Here is what was excluded and the verified reason, all checked 28 August 2026.
Continue is excluded because the project says so. Its README states the repository "is no longer actively maintained and is read-only for all users", records a final 2.0.0 release, and the continuedev organisation hosts a repo described as a "Static archive of continue.dev (acquired by Cursor)". The code stays Apache 2.0, copyright 2023 to 2026 Continue Dev, Inc., so a fork is legally clean. It is simply not something to adopt new.
Roo Code and Void are both marked archived on GitHub, at ~24k and ~29k stars, last pushed 15 May 2026 and 2 June 2026, both Apache 2.0. gpt-engineer is archived, MIT, ~55k stars, last pushed 14 May 2025. Devika is MIT and not archived, but was last pushed 25 September 2025, eleven months before we checked. SWE-agent itself, MIT at ~20k stars, is not dead but is deliberately superseded by its own maintainers, which is why the successor is ranked instead.
OpenClaw deserves a note, because at ~388k stars it is the most starred project anywhere near this space and it dominates these searches. It is real and current, its default branch was committed 28 August 2026, and its LICENSE file is the MIT licence, copyright "2026 OpenClaw Foundation", with one appended line noting that third-party notices live in THIRD_PARTY_NOTICES.md. That appended line is exactly why GitHub reports its licence as NOASSERTION rather than MIT, and why you read the file. We did not rank it because it is not a coding agent: it describes itself as a personal AI assistant for a single operator, connecting models, tools and messaging channels through one gateway.
Open weights: the models you would run behind these agents
Every agent above is a harness. Capability comes from the model, and if your reason for going open source is data control, the model decides whether code leaves your network. All licences below were read from the Hugging Face model card and LICENSE file on 28 August 2026.
Apache 2.0, unconditional. Qwen3-Coder-Next and the smaller Qwen3-Coder-30B-A3B-Instruct most teams start with on a single node, Mistral's Devstral-Small-2507, OpenAI's gpt-oss-120b and Kwaipilot's KAT-Coder-V2.5-Dev. No conditions on commercial use, no revenue trigger, no attribution requirement.
MIT. DeepSeek-V4-Pro-0813 and DeepSeek-V4-Flash-0731. Z.ai's GLM-5.2 and GLM-5.3-Flash.
Permissive with a condition, where teams get caught. Moonshot's Kimi K2 and Kimi K2.7 Code are labelled modified-mit, and the modification is explicit in the LICENSE: if the software or a derivative is used in a commercial product with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue, you must prominently display the model name in the product interface. Qwen3.8-2.4T-A95B is not Apache 2.0 at all; it ships under a bespoke "Qwen3.8-Max License" that adds a second condition on top of that trigger: if you run a Model-as-a-Service or AI work assistant business and group revenue exceeds 50 million US dollars over any twelve consecutive months, you must obtain a separate licence from Qwen before commercial use, with a carve-out for internal use not exposed to third parties. Meta's Llama family remains under its own community licence rather than an OSI-approved one, and its most recent Hugging Face release, Llama 4, was last modified in May 2025.
Filter on the licence file before you filter on the leaderboard. A condition discovered after your inference stack is built is an architecture problem, not paperwork.
What a free AI coding agent actually costs
The agent is free. The inference usually is not. Every project ranked here can be downloaded, modified and shipped inside a commercial product under MIT or Apache 2.0 with no fee, permanently. That is not the number your CFO is asking about. There are three cost shapes and you will meet some combination of all three.
Provider tokens, the default for eight of the ten. Cost scales with tokens, which scales with how much of the codebase the agent reads and how many attempts it makes, not with headcount. Two free entry points are real: Gemini CLI's documented 60 requests per minute and 1,000 per day on a personal Google account, and Codex CLI's inclusion in existing ChatGPT Plus, Pro, Business, Edu and Enterprise plans. Kilo Code's zero-markup pass-through shows you true provider cost with no vendor margin in the way.
Your own GPUs, if the point is that code never leaves your network. As a public anchor, Lambda's GPU Cloud price list on 28 August 2026 quotes on-demand H100 SXM 80 GB from 3.99 US dollars per GPU-hour and B200 SXM6 180 GB from 6.69 US dollars per GPU-hour. A single always-on H100 at that rate is roughly 2,900 US dollars a month before a useful token is served, and a model in the Qwen3-Coder-Next or DeepSeek-V4-Pro class is not a single-GPU deployment.
Engineer time, the one that gets underestimated. Somebody owns sandbox provisioning and teardown, secret scoping so no run holds a standing credential, egress policy, repository token lifetimes, upgrade cadence across projects that commit daily, model routing and failover, run logging and retention, and the on-call for all of it. Price that at fully loaded engineering cost and compare it against the token bill. For many teams the compute is the small number.
Roll it out on one queue before you roll it out on ten
What is CodeCourier? An autonomous AI software engineering platform that turns a tracked issue into a tested pull request, with every run isolated in its own disposable sandbox, agent personas that encode how your team actually writes code, a learning engine that gets sharper on your repo over time, and engineering analytics your leads can audit, across GitHub, Jira, Linear, Slack and Sentry. Talk to us about a team rollout.
Build versus buy, stated honestly
For many teams self-hosting is the right answer, and we would rather say that plainly than sell against it badly. What you get: full control of the model and the data path, provable with a network policy rather than a contract. No per-seat cost, so the economics do not punish you for hiring. No vendor, meaning no renewal negotiation, no roadmap you cannot influence, no service you cannot fork, and no supplier whose failure becomes your incident. Under an unconditional MIT or Apache 2.0 grant, today's version is yours permanently whatever the project does next. If you already run a platform team, a model gateway and an internal VM or Kubernetes fleet, the marginal work is integration rather than invention, and you should probably self-host.
What you take on. Sandbox isolation first: an agent executes untrusted, model-generated commands, so each run needs a disposable environment with no standing credentials, scoped egress and a hard time limit. That argument is made properly in why isolated code execution is non-negotiable and defined in What Is a Code Sandbox?, so we will not relitigate it. Then orchestration: queueing, retries, parallelism limits, and what happens when an agent opens a bad pull request at 02:00. Then the durable repository context that makes an agent useful past the first task, an engineering problem of its own. Then upgrades against daily-committing projects, and the on-call covering all of it. This is where open source vibe coding stops being a useful frame: prompting your way to a prototype and running an agent fleet against a shared issue tracker are different disciplines, as what agentic coding actually means sets out.
CodeCourier is the managed counterpoint. It turns a tracked issue into a tested pull request, gives every run its own isolated, disposable sandbox, encodes team conventions as agent personas rather than one generic agent, builds durable context through the learning engine, and produces analytics your leads can audit, across GitHub, Jira, Linear, Slack and Sentry. The model keys stay yours: every tier runs on your own provider keys, and you pay OpenAI, Anthropic and E2B directly while CodeCourier charges for the coordination layer. Published pricing starts at 0 US dollars for the free tier, 23 US dollars per month for Solo billed annually or 29 billed monthly, and 159 per month for Team billed annually or 199 billed monthly, with Enterprise on request. On compliance, so procurement starts from an accurate baseline: our SOC 2 page states Type I controls are mapped and the Type II observation window is open with an independent CPA firm, so the correct phrase today is in progress, not certified. GDPR and Swiss FADP posture, the DPA and the subprocessor list are on the trust centre.
Use whichever framing survives your own numbers.
Comparison table of open source AI coding agents
These projects look interchangeable in a README. Licence, isolation model and maintenance cadence are where they diverge, and those are the columns a platform review actually argues about. Every fact below appears in the entry above it.
| Agent | Licence | Language | Autonomy | Sandbox included | Bring your own model | Actively maintained |
|---|---|---|---|---|---|---|
| OpenHands | MIT | TypeScript | High | Yes, Docker, VM, own infra | Yes, plus ACP agents | Yes, 27 Aug 2026 |
| opencode | MIT | TypeScript | High | No, permissions and policies | Yes, large provider directory | Yes, 28 Aug 2026 |
| goose | Apache 2.0 | Rust | High | No, runs on the host | Yes, 15+ providers and Ollama | Yes, 27 Aug 2026 |
| Cline | Apache 2.0 | TypeScript | High | No, approvals and worktrees | Yes, keys or Cline account | Yes, 28 Aug 2026 |
| OpenAI Codex CLI | Apache 2.0 | Rust | High | Yes, sandbox and exec policy | Limited, OpenAI only | Yes, 28 Aug 2026 |
| Gemini CLI | Apache 2.0 | TypeScript | High | Yes, documented sandboxing | Limited, Google only | Yes, 26 Aug 2026 |
| Kilo Code | MIT | TypeScript | Medium to high | No, editor and CLI extension | Yes, 500+ models at cost | Yes, 28 Aug 2026 |
| mini-SWE-agent | MIT | Python | High | Yes, Docker, Podman, Apptainer | Yes, model agnostic | Yes, 17 Aug 2026 |
| Qwen Code | Apache 2.0 | TypeScript | High | Yes, per its feature table | Yes, incl. Ollama and vLLM | Yes, 28 Aug 2026 |
| Aider | Apache 2.0 | Python | Medium | No, works on your tree | Yes, almost any LLM | Slowing, 22 May 2026 |
How to run this evaluation in two weeks
Three factors decide it, none of them a feature list. First, the licence file read literally, including anything appended, because a revenue-triggered condition is an architecture constraint and not a footnote. Second, the isolation boundary: four of these ten ship one and six expect you to build it, usually the largest hidden cost in a self-hosting business case. Third, the maintenance signal with a date, weighed against whether you would genuinely maintain a fork.
The fastest honest evaluation is narrow. Take one high-volume, low-ambiguity queue, flaky test fixes, dependency bumps, small bugs with a clear reproduction, and run it two ways for two weeks: one open source agent from this list that you host, and one managed run. Measure merged pull requests, review time per pull request, and the hours your own engineers spent keeping each path alive. That last number decides it, and it is the one nobody instruments. What should not be automated on either path is the same: architectural change, ambiguous requirements and cross-service work still need a human owning the decision.
If the operational column is where your build case falls apart, bring us the queue and the numbers rather than a demo repo. Bring us your security review. Talk to us about a team rollout.
FAQ: open source AI coding agents
Is there any free AI agent for coding?
Yes, and the distinction matters: the agent is free, the inference usually is not. Every project ranked here is free to download, run and modify under MIT or Apache 2.0. What costs money is the model behind it. Gemini CLI is the closest thing to free at the point of use, with a documented tier of 60 requests per minute and 1,000 per day on a personal Google account (repo README, checked 28 August 2026). Everything else bills you as API tokens, as a seat on a vendor plan, or as GPU hours if you host open weights yourself.
Is there any open source AI for coding?
There is a lot of it, and in 2026 it is no longer second tier. Ten actively maintained projects are ranked in this article, all under OSI-approved licences: OpenHands, opencode, goose, Cline, OpenAI Codex CLI, Gemini CLI, Kilo Code, mini-SWE-agent, Qwen Code and Aider. All ten are permissively licensed for commercial use with no revenue threshold and no field-of-use restriction. The open weights you would run behind them, including Qwen3-Coder-Next and DeepSeek-V4-Pro, are open licensed too.
What is the best open source AI model for coding?
For a team that wants zero licence conditions, Qwen3-Coder-Next under Apache 2.0 and DeepSeek-V4-Pro-0813 under MIT are the two cleanest choices (Hugging Face model cards, checked 28 August 2026). If you can accept a conditional licence, Kimi K2.7 Code is modified MIT: you must display the model name once your product passes 100 million monthly active users or 20 million US dollars in monthly revenue. Read the licence file, not the licence tag, before you commit.
Is OpenClaw free?
Yes. OpenClaw ships under the MIT licence, with a closing line pointing to THIRD_PARTY_NOTICES.md for incorporated code (LICENSE file, checked 28 August 2026). It is free to run, and at ~388k stars it is the most starred project anywhere near this space. It is not a team coding agent, though: it describes itself as a personal AI assistant designed for a single operator, connecting models, tools and messaging channels through one gateway. It is not what you evaluate for a CI-attached engineering workflow.
What is the best open source coding agent for a platform team?
OpenHands, if you want a self-hosted control plane rather than a single terminal binary. It is MIT licensed, it runs agents across Docker containers, VMs or your own company infrastructure, and it drives third-party agents through the Agent Client Protocol rather than locking you to one. If vendor-neutral governance matters more, goose is the only project here with a foundation behind it: it is part of the Agentic AI Foundation at the Linux Foundation.
How do I run a self-hosted AI coding assistant securely?
Assume the agent will execute untrusted, model-generated commands, because that is exactly what it does. Give every run a disposable environment with no standing credentials, scoped network egress, a short-lived repository token and a hard wall clock limit, then destroy it. Of the ten projects ranked here, four ship execution isolation you can turn on: OpenHands, OpenAI Codex CLI, Gemini CLI and mini-SWE-agent. For the other six, that boundary is something your platform team builds and operates.