Citation and evidence

Best Open-Source Coding Agents

34 min full readUpdated 88 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI AgentsAI Code GenerationDeveloper ToolsOpen Source AI

Cite this article

This page compares open-source coding agents: AI coding tools whose own source code is published under a licence approved by the Open Source Initiative (OSI), so that anyone can read, modify and redistribute the agent itself. As of 19 November 2026 it profiles thirteen actively released agents that passed the inclusion criteria set out below: opencode, Codex CLI, Pi, Gemini CLI, OpenHands, Cline, goose, Aider, Qwen Code, Kilo Code, mini-swe-agent, Kimi Code CLI and Mistral Vibe. All thirteen use the MIT License or the Apache License 2.0, both of which are on the OSI's list of approved licences. [1] The tables below order them by GitHub stars on that date, which is a count of bookmarks and not a measure of quality. The page does not rank the tools or recommend one of them.

This page is about the software that drives a model (the harness), not about the models. "Open source" here describes the agent's code only. The models an agent calls are a separate question: a model with downloadable weights is called open-weight on this wiki, and an open-source agent can just as well be pointed at a closed, hosted model. Several well-known agents publish some code or a public repository without an open-source licence; they are listed separately under Not open source with the licence they actually use. For a comparison that includes proprietary tools and prices, see Best AI Coding Assistants; for background on what a coding agent is, see that article.

How the tools were selected

Nothing on this page is an editorial ranking. The criteria below are this article's own inclusion rules, not an external standard. A tool is in the main table only if it passed three checks, all made on 19 November 2026:

  1. Licence. The LICENSE file in the project's current repository is an OSI-approved licence that covers the agent itself. [1] Repository moves were followed: sst/opencode now resolves to anomalyco/opencode, block/goose to aaif-goose/goose, All-Hands-AI/OpenHands to OpenHands/OpenHands, and badlogic/pi-mono to earendil-works/pi. [2][4][6][8]
  2. Maintenance. The repository is not archived, its README does not declare the project shut down, unmaintained or superseded, and a release was published in the twelve months before 19 November 2026.
  3. Adoption. The repository has at least 5,000 GitHub stars. Stars are used only as a documented threshold of public interest.

The main table is limited to the projects reviewed for this article. It is not a census of every project that would pass the three checks. Two projects that passed them on the same date but are not profiled here are Open Interpreter, whose README calls it "a fork of OpenAI's Codex" (Apache-2.0; release rust-v0.0.61 on 11 November 2026; 68,910 stars), and Forge (tailcallhq/forgecode; Apache-2.0; v2.14.0 on 1 October 2026; 7,635 stars). [83][84] Freebuff (CodebuffAI/freebuff; Apache-2.0; 14,233 stars) publishes no GitHub releases and was not assessed against the maintenance check. [85]

Projects that pass the licence check but fail the maintenance check are in Archived, superseded or unmaintained projects. Editors and terminals that contain an agent are in Editors and terminals with a built-in agent. Benchmark figures are quoted only from the benchmark's own leaderboard, with the date, the model and the harness version where the source gives one. None of these tools was installed or run for this article; every statement comes from the project's repository, its documentation or a published advisory.

Comparison table

Licence, release and star figures come from each repository and the GitHub API on 19 November 2026. "Latest release" is the newest non-prerelease tag for the agent itself.

ToolLicenceMaintainerInterfaceLatest releaseGitHub stars (2026-11-19)
opencodeMITAnomaly (anomalyco)Terminal, desktop app (beta)v2.3.2, 2026-11-18 (1.x line: v1.19.6, 2026-11-17)227,832 [2]
Codex CLIApache-2.0OpenAITerminal0.170.0, 2026-11-18135,798 [3]
PiMITEarendil Inc.Terminal, RPC, TypeScript SDKv1.5.2, 2026-11-10122,758 [4]
Gemini CLIApache-2.0GoogleTerminalv0.68.0, 2026-11-10108,925 [5]
OpenHandsMITOpenHandsSelf-hosted web UI (Agent Canvas), SDKv1.42.0, 2026-11-1893,768 [6]
ClineApache-2.0Cline Bot Inc.VS Code extension, CLI, desktop app, SDKv4.1.26, 2026-10-2671,758 [7]
gooseApache-2.0Agentic AI Foundation (Linux Foundation)Desktop app, CLI, APIv1.60.1, 2026-11-1856,765 [8]
AiderApache-2.0Aider-AITerminal0.86.2 on PyPI, 2026-02-1250,171 [9][15]
Qwen CodeApache-2.0Qwen team (QwenLM)Terminal, desktop app, Web UI (experimental), IDE plugins, SDKsv0.28.0, 2026-11-1730,764 [10]
Kilo CodeMITKilo Code (Anaconda announced its acquisition on 15 July 2026)VS Code extension, JetBrains plugin, CLIv7.8.27, 2026-11-0528,052 [11][66]
mini-swe-agentMITPrinceton and Stanford researchersTerminal, Python libraryv2.4.6, 2026-07-239,759 [12]
Kimi Code CLIMITMoonshot AITerminal2.6.3, 2026-11-178,253 [13]
Mistral VibeApache-2.0Mistral AITerminalv3.1.3, 2026-11-185,424 [14]

Expanded article table

Five rows need a caveat. Aider's newest GitHub release is v0.86.0 (9 August 2025), its newest PyPI release is 0.86.2 (12 February 2026), and the last commit on its default branch was on 22 May 2026, so it is the least recently updated tool in the table. [9][15] The OpenHands/OpenHands repository now holds Agent Canvas, a control centre that runs the OpenHands agent and third-party agents; the agent, tools and server live in OpenHands/software-agent-sdk, also under the MIT License (v1.62.0, 11 November 2026). [6][16] Cline's README states that its JetBrains plugin is not open-sourced, so the Apache-2.0 licence covers the SDK, CLI, VS Code extension and desktop app but not that plugin. [7] Kilo's MIT licence covers the Kilo-Org/kilocode repository, which holds the VS Code extension, the JetBrains plugin and the CLI; Kilo's cloud platform and its desktop app are in separate repositories, Kilo-Org/cloud and Kilo-Org/kilo-desktop, under the Open Core Ventures Source Available License version 1.0, which is not on the OSI's list. [1][11][81][82] The opencode repository also bundles Figma skill files that are third-party material; their notice says "OpenCode's MIT license does not relicense them". [88]

What each tool is

ToolSummary from the current README
opencodeTerminal agent with two built-in agents, build (full access) and plan (read-only, denies edits by default), plus a general subagent [2]
Codex CLIOpenAI's terminal agent; sign in with a ChatGPT plan or use an API key. The README points to a separate IDE extension, desktop app and cloud agent [3]
PiA "minimal, extensible agent harness" customised through extensions, skills, prompt templates and themes; it leaves out sub-agents and plan mode by design [4]
Gemini CLIMaintained for Gemini Code Assist Standard and Enterprise licences; free individual access and the Google AI Pro and Ultra plans moved to the Antigravity CLI, which is not open source, on 18 June 2026 [5][34]
OpenHandsRuns agents locally, in Docker containers, on virtual machines or in the vendor's cloud, with scheduled and webhook-triggered automations [6]
ClinePlan and Act modes, checkpoints, .clinerules files, Model Context Protocol servers, multi-agent teams, scheduled agents and a headless mode for CI [7]
gooseGeneral-purpose local agent written in Rust with more than 15 model providers and more than 70 MCP extensions [8]
AiderPair programmer that builds a map of the repository and commits each change to git with a generated message [9]
Qwen CodeMulti-protocol agent that began as a fork of Gemini CLI v0.8.2 and stopped syncing with it from Qwen Code v0.1 [10]
Kilo CodeCode, Plan, Ask and Debug agents, inline autocomplete and a marketplace; its CLI is a fork of opencode [11]
mini-swe-agentAn agent class of about 100 lines of Python whose only tool is bash; the harness behind the SWE-bench "Bash Only" leaderboard [12][38]
Kimi Code CLISingle-binary terminal agent with video input that speaks the Agent Client Protocol to editors [13]
Mistral VibeTerminal agent that asks before risky actions; the project takes bug reports but does not accept pull requests [14]

Expanded article table

Editors and terminals with a built-in agent

Two open-source developer tools are larger than a coding agent but ship one. They are listed apart from the main table because the agent is one feature of a larger program.

ToolLicenceMaintainerAgentLatest releaseGitHub stars (2026-11-19)
ZedGPL-3.0-or-later, with Apache-2.0 componentsZed IndustriesFirst-party agent in the Agent Panel; external agents connect through the Agent Client Protocol [78]v1.29.2, 2026-11-1895,289 [17]
WarpAGPL-3.0, with the UI framework crates under MITWarpBuilt-in coding agent; can also host other CLI agentsv0.2026.06.03.09.49.stable_00, 2026-06-03 (newest stable release published on GitHub; the repository was last pushed on 2026-11-18)66,058 [18]

Expanded article table

Zed's README says its source is "licensed primarily under GPL-3.0-or-later, with Apache-2.0 components where marked". [17] Warp announced on 28 April 2026 that its client was now open source; the repository's licensing section puts the warpui_core and warpui crates under MIT and the rest under the AGPL v3. [18][19] Both the GNU General Public License version 3 and the GNU Affero General Public License version 3 are on the OSI's list of approved licences. [1] How their conditions differ from those of MIT and Apache-2.0 is explained in copyleft.

Archived, superseded or unmaintained projects

These projects carry an OSI-approved licence, so the code can still be forked, but each fails the maintenance check: the repository is archived, the README declares the project unmaintained or superseded, or no release was published in the twelve months before 19 November 2026.

ProjectLicenceStatus on 2026-11-19Last releaseStars
Roo CodeApache-2.0Repository archived; README says the extension "was shut down on May 15th" and points to the community fork ZooCode and to Clinev3.54.0, 2026-05-1524,238 [20][21]
ContinueApache-2.0README: the repository "is no longer actively maintained and is read-only for all users"2.0.0 (tag v2.0.0-vscode, 2026-06-19)36,512 [23]
SWE-agentMITREADME: "Most of our current development effort is on mini-swe-agent, which has superseded SWE-agent"v1.1.0, 2025-05-2220,829 [25]
PlandexMITNo release in the previous twelve months; last commit on the default branch 2025-10-03; README says Plandex Cloud is winding downcli/v2.2.1, 2025-07-1615,923 [26]
VoidApache-2.0Repository archived; README: "Void is deprecated and no longer accepting contributions"v1.3.4, 2025-04-15 (marked as a pre-release)28,707 [27]
OpenCode (Go, opencode-ai/opencode)MITArchived; README says the project "has continued under the name Crush"v0.0.55, 2025-06-2713,917 [28]
Kimi CLI (Python)Apache-2.0Archived; description points to Kimi Code CLI1.52.0, 2026-09-2211,410 [29]

Expanded article table

ZooCode, the fork that the Roo Code notice recommends, is published at Zoo-Code-Org/Zoo-Code under Apache-2.0; its latest release was v3.98.0 on 15 November 2026 and it had 2,135 stars, below this page's adoption threshold. [22]

Not open source

The tools below are often discussed alongside open-source agents. Each has a public repository or a published component, but the licence is not OSI-approved, or the repository contains no source code. The right-hand column quotes or summarises the licence file or README and adds nothing to it.

ToolLicence as publishedWhat the licence file or README says
Crush (Charm)Functional Source License, Version 1.1, MIT Future License (FSL-1.1-MIT)"A Permitted Purpose is any purpose other than a Competing Use." The licence also grants "an additional license to use the Software under the MIT license that is effective on the second anniversary of the date we make the Software available" [30]
Claude CodeLICENSE.md, in full: "© Anthropic PBC. All rights reserved. Use is subject to Anthropic's Commercial Terms of Service."The repository has no other licence file [32]
GitHub Copilot CLIGitHub Copilot CLI LicenseGrants a licence "to install and run copies of the GitHub Copilot CLI" and to "reproduce and redistribute unmodified copies of the Software as part of an application or service", not on a standalone basis; it does not grant the right to "Modify, adapt, translate, or create derivative works of the Software" [33]
Antigravity CLI (Google)No licence fileThe repository holds a README, a changelog and example scripts, not the program's source; the README links Google's terms of service for the product [34]
TabbyApache 2.0, except the ee/ directory, which is under "The Tabby Enterprise license"The root LICENSE says content under ee/ is licensed under the licence defined in ee/LICENSE; that file says the software "may only be used in production" under TabbyML's subscription terms with a valid Tabby Enterprise licence [35][87]
Cline JetBrains pluginSource not publishedThe Cline README's product index says: "Currently we are not open-sourcing JetBrains plugins" [7]
Kilo cloud platform and Kilo DesktopOpen Core Ventures Source Available License (OCVSAL) version 1.0"Using software and associated documentation files (the "Software") in production requires a valid commercial agreement from the copyright holder." [81][82]

Expanded article table

The Functional Source License is not on the OSI's list of approved licences, and its own site describes it as "a Fair Source license that converts to Apache 2.0 or MIT after two years". [1][31] The archived Go OpenCode repository, whose README says the project continued as Crush, is under the MIT License; Crush's licence file is FSL-1.1-MIT. [28][30] Licences that publish source code without meeting the OSI's definition are covered in source-available. The arrangement in which a company publishes a core under an open-source licence and keeps other parts closed is described in open core; the general differences between licence families are covered in open-source license.

Published benchmark results

Benchmark scores for coding agents measure an agent and a model together, and each leaderboard uses its own tasks, time limits and model versions. Figures from different leaderboards, different versions of one leaderboard, or different models are not comparable with each other, and none of the numbers below should be read as a ranking of the tools on this page.

Terminal-Bench 4.1

Terminal-Bench scores agent and model pairs on command-line tasks. Version 4.1 was published on 9 November 2026 with hardened task verifiers; its maintainers re-graded the 4.0 submissions and say all scores stayed within the previously reported 95% confidence intervals. [37] On 19 November 2026 the public 4.1 leaderboard listed 52 entries for seven agents: GitHub Copilot, Claude Code, Antigravity CLI, Codex, Muse Code, Grok Build and mini-swe-agent (two of its eleven entries are spelled "mini-SWE-agent"). [36] Two of those names correspond to tools in this page's main table. The board's "Codex" entries link to OpenAI's Codex product page and do not say which Codex client or version was run; the mini-swe-agent entries link to that project's repository or website. [36] The rows below are the two highest entries under each of the two names:

Agent as listedModel (effort)Resolution rate (± 95% confidence interval)Release date shownSource
CodexGPT-6.2 Astra (max)62.1% ± 2.32026-11-17[36]
CodexGPT-6.2 Sol (max)61.5% ± 2.82026-11-17[36]
mini-swe-agentMiMo V2.6 Pro (none)27.0% ± 3.62026-09-22[36]
mini-swe-agentEmber-1 (none)21.8% ± 3.52026-09-24[36]

Expanded article table

The leaderboard table names the agent but not its version, and the highest entry overall on that date was GitHub Copilot with Copilot Fusion (max) at 71.2% ± 2.9. [36] The Codex and mini-swe-agent rows use different models, so the gap between them says nothing about the two harnesses. The other eleven tools in the main table had no entry. Gemini CLI's own README tells readers not to use it to benchmark Gemini models, because its harness is not tuned for the Gemini 4 series. [5]

SWE-bench Verified

SWE-bench Verified is a 500-task, human-filtered subset of SWE-bench, which is built from real GitHub issues. The official site keeps a view of submissions from any agent and a "Bash Only" view, which it describes as "The default Verified view: every model in the same mini-SWE-agent environment". [38] The newest Verified entry in the site's data on 19 November 2026 was dated 26 February 2026, so these results are older than the releases in the comparison table.

AgentModel% resolvedEntry dateNote
mini-swe-agent 2.0.0Claude 4.5 Opus (high)76.82026-02-17Bash Only view; the highest mini-swe-agent entry [38]
mini-swe-agent 2.0.0MiniMax M2.5 (high)75.82026-02-17Bash Only view; marked as an open-weights model; Gemini 3 Flash (high) has the same score [38]
OpenHandsGPT-571.82025-08-07Marked as performed or checked by the SWE-bench team [38]
OpenHandsClaude 4 Sonnet70.42025-05-24Marked as performed or checked by the SWE-bench team [38]
SWE-agentClaude 4 Sonnet66.62025-05-22Marked as performed or checked by the SWE-bench team [38]

Expanded article table

The rows differ in model, harness version and date, so they do not compare the three agents with each other. The mini-swe-agent README separately claims more than 74% on SWE-bench Verified without naming the model. [12]

Running an open stack end to end

A reader who wants every layer to be inspectable needs an open-source agent, an open-weight model and a local inference server. In the documentation read for this article the agent reaches the server over HTTP, usually through an OpenAI-compatible API: opencode's documentation, for example, reaches Ollama, LM Studio and vLLM at /v1 endpoints, and Pi's documents the llama.cpp router server. [44][46] Setting up the server side is covered in how to run LLMs locally, self-hosting LLMs and local AI; model choices by memory size are in Best Local and On-Device LLMs and Best Open-Source LLMs.

Which agents document local models

ToolLocal runtimes named in the documentationDocumented requirement or caveat
opencodeOllama, LM Studio and vLLM with built-in discovery; any OpenAI-compatible endpoint as a custom providerOnly Ollama models whose metadata includes the completion capability are listed; discovered vLLM models start with tools disabled [44]
Codex CLIollama and lmstudio are built-in provider IDs; custom providers take a base URLProvider settings must be in the user-level configuration file; project files cannot change them [41]
Pillama.cpp router server; custom providersThe router loads and unloads GGUF models on demand [46]
OpenHandsLM Studio, Ollama, vLLM, SGLang"Local LLMs can have limited functionality"; a GPU-backed server is advised [45]
ClineOllama, LM Studio, Atomic Chat; any OpenAI-compatible APIEnable "Use Compact Prompt"; 16-32 GB of RAM for small quantized models, 32-64 GB for mid-size coding models, 64 GB or more for larger models and context windows [7][40]
gooseOllama, LM Studio; OpenAI-compatible endpoints such as vLLM"goose relies heavily on tool calling capabilities"; a model without function calling can only do chat completion, with all extensions disabled [42]
AiderOllama and other local serversOllama's default 2k context window is too small and silently discards excess context; the guide starts the server with an 8k window [39]
Qwen CodeOllama, vLLM; OpenAI, Anthropic and Gemini protocolsNone stated in the README [10]
Kilo CodeOllamaLocal models are more likely to loop or call tools incorrectly; the guide recommended qwen3-coder:30b when read [43]
mini-swe-agentAny model through LiteLLM; a vLLM example for SWE-bench runsVersion 2 calls its bash tool through the model's native tool-calling interface by default; a text-based mode that parses commands from the reply remains for models without one [12][49][80]
Kimi Code CLIThird-party services through the OpenAI-compatible protocolWorks out of the box with Moonshot's Kimi models [13][48]

Expanded article table

Gemini CLI is absent from this table because its README describes access only through Google licences and keys, and Mistral Vibe's README describes signing in with a Mistral account or API key. [5][14] Zed lists Ollama, LM Studio and OpenAI-compatible endpoints among the providers for its own agent. [47]

The caveats in the table come from the projects themselves, and they fall into three groups: context size (Aider), tool calling (goose), and the lower reliability of small local models (Kilo Code and OpenHands). [39][42][43][45] The Terminal-Bench 4.1 board gives a dated data point on open-weight models in agent harnesses. The best mini-swe-agent entry on 19 November 2026 was MiMo V2.6 Pro at 27.0%, while Kimi K3.1 Pro scored 59.7% ± 4.2 inside the proprietary Claude Code; this wiki lists both models as open-weight. [36] The two entries differ in both model and harness, so they do not isolate the effect of either.

Telemetry and data handling

An open-source licence does not mean an agent sends nothing home. The defaults below are as each project documents them; the table covers only projects whose documentation on this point was read for this article. Whatever the agent's own telemetry setting, prompts and code are still sent to whichever model provider is configured unless that provider is a local server.

ToolDocumented default
AiderAnalytics are collected only after the user opts in; the docs say code, chat messages and keys are never collected [50]
ClineUsage and error telemetry is on by default in the extensions, desktop app and CLI, and can be turned off in settings or with telemetryOptOut [51]
gooseAsks on first use for permission to collect anonymous usage data [52]
PienableInstallTelemetry (anonymous install and update reporting) defaults to true; enableAnalytics defaults to false [53]
Gemini CLICollects usage statistics; the terms page says users may opt out of sending them to Google [54]
Mistral VibeTelemetry and crash reporting are turned off by setting enable_telemetry = false [14]
opencodeThe Enterprise page of the V1 documentation says: "OpenCode does not store your code or context data. All processing happens locally or through direct API calls to your AI provider." The V2 documentation read for this article has no page on telemetry or data handling [55]
ZedCollects anonymous telemetry; client-side metrics and crash reports can be disabled in settings [56]
Continue (final release)The 2.0.0 release removed anonymous telemetry [23]

Expanded article table

Sandboxing and permission models

What an agent may do without asking differs from tool to tool.

ToolDocumented model
Codex CLICommands run inside an operating-system sandbox by default: Seatbelt on macOS, bubblewrap on Linux and WSL2, a native sandbox on Windows. Sandbox modes include a read-only mode; approval policies include on-request and never [57]
Gemini CLIDocuments five sandboxing methods: macOS Seatbelt, Docker or Podman containers, a Windows native sandbox, gVisor and LXC/LXD (experimental) [58]
OpenHandsCan run without a sandbox (the README warns that the agent then has full access to the filesystem), in one Docker container, or with a separate container per conversation [6]
mini-swe-agentRuns each action as an independent subprocess; supports local, Docker or Podman, Singularity or Apptainer, and bubblewrap environments [12]
Qwen CodeBuilt-in sandboxing and optional git worktrees for wide-ranging edits [10]
ClineEvery file edit and terminal command needs approval unless auto-approve is switched on [7]
opencodeRules of allow, deny or ask per action and resource. The base policy allows every action and asks before reading .env files [60]
goose"Autonomous Mode is applied by default"; a manual-approval mode and a risk-based approval mode can be selected [59]
Kilo Codekilo run --auto approves permission prompts automatically unless a rule denies the action; the README says to use it only in trusted environments [11]
Pi"Pi does not include a built-in permission system"; the project advises running it in a container, virtual machine or sandbox [4][61]
Zedagent.tool_permissions with a global default and per-tool always_allow, always_deny and always_confirm patterns [62]

Expanded article table

Licence changes and governance events

A licence file describes the code as it stands today. Who maintains a project, and under what terms new versions appear, can change. The events below are from primary sources unless marked.

DateEventSource
2025-12-09The Linux Foundation announces the Agentic AI Foundation, with Block contributing goose alongside Anthropic's Model Context Protocol and OpenAI's AGENTS.md[63]
2026-02-22Kilo-Org/kilocode-legacy is created to hold Kilo's earlier extension under Apache-2.0; it is now archived, and its README says the legacy extensions reached end of life on 31 July 2026. The current Kilo-Org/kilocode repository is MIT, with copyright lines for Kilo Code and opencode[11][64]
2026-03-30Earendil's licensing RFC for Pi states that "Pi remains MIT licensed and that will not change", while future related parts may be Fair Source or proprietary[68]
2026-04-08Earendil Inc. announces its acquisition of the Pi project from its author, Mario Zechner[67]
2026-04-21Kilo's blog reports that Roo Code is shutting down and notes that "Kilo started as a fork of Roo"[65]
2026-04-28Warp opens the source of its client[19]
2026-05-15The Roo Code extension is shut down; the repository, since archived, publishes its last release (v3.54.0) the same day[20][21]
2026-06-18Free individual access to Gemini CLI ends; Google directs those users to Antigravity CLI, which is not open source[5][34]
2026-06-18The trade site Web Developer (a secondary source) gives this as the date on which Cursor announced that it had acquired Continue. Continue's own website says "Continue was acquired by Cursor" without giving a date. The final release tag, v2.0.0-vscode, was published on GitHub on 19 June 2026 (UTC), and the README calls the repository read-only[23][24][79]
2026-07-15Anaconda's blog announces the acquisition of Kilo Code, referring to Kilo's "open source and source-available codebase"[66]

Expanded article table

Several lineages connect tools on this page. Roo Code began as a fork of Cline (its shutdown notice calls Cline the project "from where Roo Code originated"), and Kilo began as a fork of Roo. [21][65] Kilo's current CLI is, in its README's words, "a fork of OpenCode", meaning Anomaly's opencode. [11] Qwen Code was originally based on Gemini CLI. [10] Kimi Code CLI's README says its terminal interface is built on Pi's pi-tui library. [13] The archived Roo Code and Continue repositories remain available under Apache-2.0. [20][23]

Choosing by documented capability

These pointers restate facts from the tables above; they are not recommendations.

  • Operating-system sandbox on by default: Codex CLI's documentation says commands run inside an OS-level sandbox by default. [57]
  • Approval before every edit and command by default: Cline. [7]
  • Editor integration: Cline (VS Code, with a closed JetBrains plugin) and Kilo Code (VS Code and JetBrains) ship editor extensions; Kimi Code CLI and other agents can be driven from Zed and other editors over the Agent Client Protocol. [7][11][13][78]
  • Self-hosted web interface and automations: OpenHands. [6]
  • Foundation governance, not a single vendor: goose, a project of the Agentic AI Foundation. [8][63]
  • Local servers detected without manual configuration: opencode documents built-in discovery for Ollama, LM Studio and vLLM. [44]
  • Bash as the only tool: mini-swe-agent. Version 2 calls it through the model's native tool-calling interface by default and keeps a text-based mode for models without one. [12][80]
  • Analytics off unless the user opts in: Aider. [50]
  • Small core by design: mini-swe-agent (its README describes an agent class of about 100 lines) and Pi (described by its maintainers as minimal). [4][12]

Known limits and security issues

Coding agents read untrusted text (source files, issue comments, web pages, tool output) and can run commands, which makes them a target for prompt injection. The OWASP Gen AI Security Project describes indirect prompt injection as occurring when a model accepts input from external sources such as websites or files whose content alters its behaviour. [69] A 2025 study by Liu and colleagues built a test framework, AIShellJack, with 314 attack payloads covering 70 MITRE ATT&CK techniques, ran it against GitHub Copilot and Cursor, and reported attack success rates as high as 84% for executing malicious commands. [70] Pi's security guide makes the same point about its own tool: files, comments, command output and model responses can steer the model, and reviewing the transcript does not create a security boundary. [61]

Open-source agents have also had conventional software vulnerabilities, published as GitHub security advisories. The table gives one advisory per project as an example. It is not a list of each project's most severe issue.

ProjectAdvisoryPublishedSummaryAffected and fixed versions
ClineGHSA-9ppg-jx86-fqw72026-02-17A compromised npm token was used to publish cline@2.3.0 with a postinstall script that installed an unrelated package; a corrected 2.4.0 was published about eight hours later. The VS Code extension was not affected [71]cline 2.3.0 on npm; fixed in 2.4.0
opencodeCVE-2026-228122026-01-12Unauthenticated HTTP server allowed arbitrary command execution [72]opencode-ai before 1.0.216
Codex CLICVE-2025-595322025-09-19Sandbox bypass caused by a bug in path configuration logic [73]@openai/codex 0.2.0 to 0.38.0; fixed in 0.39.0
ZedCVE-2026-279672026-02-25Symlink escape in the agent's file tools [74]Before 0.225.9
gooseCVE-2026-727182026-07-24Arbitrary command execution in goose review through a git core.fsmonitor setting [75]Before 1.44.0
OpenHandsCVE-2026-337182026-03-23Command injection in the git diff handler [76]1.4.0 and earlier; fixed in 1.5.0
PiCVE-2026-543252026-06-08Project-local extensions were loaded without approval [77]Before 0.79.0

Expanded article table

On 19 November 2026 the repositories listed 12 published advisories for Zed, 4 for Pi, 3 each for Cline and opencode, and 1 each for Codex, goose and OpenHands; Cline, opencode and Zed each had one rated critical. [2][3][4][6][7][8][17] OpenHands's software-agent-sdk repository listed two more, both published on 13 November 2026, in which the OpenAI-compatible gateway let agent commands run outside the Docker sandbox in openhands-agent-server 1.49.0 to 1.60.2 (fixed in 1.61.0). [86]

Two of the entries in the table concern the safeguards themselves (a sandbox bypass in Codex CLI and a file-tool escape in Zed), and the Cline entry concerns the package registry through which the tool is distributed. The table is not a measure of which tool is less safe: Aider, Gemini CLI, Kilo Code, Qwen Code, mini-swe-agent, Kimi Code CLI and Mistral Vibe had no published advisories on GitHub on 19 November 2026, which does not show that they have had fewer flaws. [5][9][10][11][12][13][14] Among the mitigations the projects document, Pi's guide advises running the agent inside a container, virtual machine or sandbox and limiting the credentials and network services available to it, and Codex's documentation pairs its sandbox with an approval policy under which the agent asks before going beyond the sandbox boundary. [57][61]

Continue reading

More to read

Topics:AI AgentsAI Code GenerationDeveloper ToolsOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5Open Source Initiative. "OSI Approved Licenses". opensource.org, accessed 19 November 2026. opensource.org/licenses
  2. ^1 ^2 ^3 ^4anomalyco. "opencode" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...opencode
  3. ^1 ^2 ^3OpenAI. "codex" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...codex
  4. ^1 ^2 ^3 ^4 ^5 ^6earendil-works. "pi" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...pi
  5. ^1 ^2 ^3 ^4 ^5 ^6google-gemini. "gemini-cli" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...gemini-cli
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7OpenHands. "OpenHands" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...OpenHands
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9cline. "cline" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...cline
  8. ^1 ^2 ^3 ^4 ^5aaif-goose. "goose" (repository, README, LICENSE and security advisories). GitHub, read 19 November 2026. github.com/...goose
  9. ^1 ^2 ^3 ^4Aider-AI. "aider" (repository, README and LICENSE.txt). GitHub, read 19 November 2026. github.com/...aider
  10. ^1 ^2 ^3 ^4 ^5 ^6QwenLM. "qwen-code" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...qwen-code
  11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Kilo-Org. "kilocode" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...kilocode
  12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8SWE-agent. "mini-swe-agent" (repository, README and LICENSE.md). GitHub, read 19 November 2026. github.com/...mini-swe-agent
  13. ^1 ^2 ^3 ^4 ^5 ^6MoonshotAI. "kimi-code" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...kimi-code
  14. ^1 ^2 ^3 ^4 ^5mistralai. "mistral-vibe" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...mistral-vibe
  15. ^1 ^2Python Package Index. "aider-chat" (release history). pypi.org, read 19 November 2026. pypi.org/...aider-chat
  16. ^OpenHands. "software-agent-sdk" (repository and LICENSE). GitHub, read 19 November 2026. github.com/...software-agent-sdk
  17. ^1 ^2 ^3zed-industries. "zed" (repository, README, licence files and security advisories). GitHub, read 19 November 2026. github.com/...zed
  18. ^1 ^2warpdotdev. "warp" (repository, README and licence files). GitHub, read 19 November 2026. github.com/...warp
  19. ^1 ^2Zach Lloyd. "Warp is now open-source". Warp, 28 April 2026. warp.dev/...warp-is-now-open-source
  20. ^1 ^2 ^3RooCodeInc. "Roo-Code" (archived repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...Roo-Code
  21. ^1 ^2 ^3Roo Code. "Roo Code Docs" (Extension Shutdown notice, last updated 15 May 2026). roocodeinc.github.io/Roo-Code
  22. ^Zoo-Code-Org. "Zoo-Code" (repository and LICENSE). GitHub, read 19 November 2026. github.com/...Zoo-Code
  23. ^1 ^2 ^3 ^4continuedev. "continue" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...continue
  24. ^"Cursor Acquires Continue, Putting a Pioneering Open-Source Coding Agent Under Commercial Control". Web Developer, 18 June 2026. webdeveloper.com/...res-continue-open-source-agent
  25. ^SWE-agent. "SWE-agent" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...SWE-agent
  26. ^plandex-ai. "plandex" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...plandex
  27. ^voideditor. "void" (archived repository, README and LICENSE.txt). GitHub, read 19 November 2026. github.com/...void
  28. ^1 ^2opencode-ai. "opencode" (archived repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...opencode
  29. ^MoonshotAI. "kimi-cli" (archived repository). GitHub, read 19 November 2026. github.com/...kimi-cli
  30. ^1 ^2charmbracelet. "crush" (repository, README and LICENSE.md). GitHub, read 19 November 2026. github.com/...crush
  31. ^"FSL - Functional Source License". fsl.software, accessed 19 November 2026. fsl.software
  32. ^Anthropic. "claude-code" LICENSE.md. GitHub, read 19 November 2026. github.com/...LICENSE.md
  33. ^GitHub. "copilot-cli" LICENSE.md (GitHub Copilot CLI License). GitHub, read 19 November 2026. github.com/...LICENSE.md
  34. ^1 ^2 ^3google-antigravity. "antigravity-cli" (repository and README). GitHub, read 19 November 2026. github.com/...antigravity-cli
  35. ^TabbyML. "tabby" LICENSE. GitHub, read 19 November 2026. github.com/...LICENSE
  36. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Terminal-Bench. "Terminal-Bench 4.1" leaderboard. tbench.ai, accessed 19 November 2026. tbench.ai
  37. ^Terminal-Bench. "Terminal-Bench 4.1". Terminal-Bench news, 9 November 2026. tbench.ai/...terminal-bench-4-1
  38. ^1 ^2 ^3 ^4 ^5 ^6 ^7SWE-bench. "SWE-bench Leaderboards". swebench.com, accessed 19 November 2026. swebench.com
  39. ^1 ^2Aider. "Ollama". Aider documentation, accessed 19 November 2026. aider.chat/...ollama
  40. ^Cline. "Local models". Cline documentation, accessed 19 November 2026. docs.cline.bot/...overview
  41. ^OpenAI. "Advanced Configuration". Codex documentation, accessed 19 November 2026. learn.chatgpt.com/...config-advanced
  42. ^1 ^2goose. "Configure LLM Provider" (Supported LLM Providers). goose documentation, accessed 19 November 2026. goose-docs.ai/...providers
  43. ^1 ^2Kilo Code. "Using Ollama with Kilo Code". Kilo documentation, accessed 19 November 2026. kilo.ai/...ollama
  44. ^1 ^2 ^3OpenCode. "Providers". OpenCode documentation, accessed 19 November 2026. opencode.ai/...providers
  45. ^1 ^2OpenHands. "Run Local LLMs with OpenHands". OpenHands Docs, accessed 19 November 2026. docs.openhands.dev/...local-llms
  46. ^1 ^2earendil-works. "Local Models with llama.cpp". Pi coding agent documentation, read 19 November 2026. github.com/...llama-cpp.md
  47. ^Zed Industries. "Use API Access". Zed documentation, accessed 19 November 2026. zed.dev/...use-api-access
  48. ^Moonshot AI. "Providers and models". Kimi Code CLI Docs, accessed 19 November 2026. moonshotai.github.io/...providers
  49. ^SWE-agent. "Local models". mini-SWE-agent documentation, accessed 19 November 2026. mini-swe-agent.com/...local_models
  50. ^1 ^2Aider. "Analytics". Aider documentation, accessed 19 November 2026. aider.chat/...analytics
  51. ^Cline. "Cline Telemetry". Cline documentation, accessed 19 November 2026. docs.cline.bot/...telemetry
  52. ^goose. "Anonymous Usage Data". goose documentation, accessed 19 November 2026. goose-docs.ai/...usage-data
  53. ^earendil-works. "Settings Reference" (Updates, telemetry, and warnings). Pi coding agent documentation, read 19 November 2026. github.com/...settings.md
  54. ^Google. "Gemini CLI: License, Terms of Service, and Privacy Notices" (last updated 10 April 2026). Gemini CLI documentation. geminicli.com/...tos-privacy
  55. ^OpenCode. "Enterprise". OpenCode V1 documentation (a previous version of the docs), accessed 19 November 2026. opencode.ai/...enterprise
  56. ^Zed Industries. "Telemetry in Zed". Zed documentation, accessed 19 November 2026. zed.dev/...telemetry
  57. ^1 ^2 ^3OpenAI. "Sandbox". Codex documentation, accessed 19 November 2026. learn.chatgpt.com/...sandboxing
  58. ^Google. "Sandboxing in Gemini CLI". Gemini CLI documentation, accessed 19 November 2026. geminicli.com/...sandbox
  59. ^goose. "goose Permission Modes". goose documentation, accessed 19 November 2026. goose-docs.ai/...goose-permissions
  60. ^OpenCode. "Permissions". OpenCode documentation, accessed 19 November 2026. opencode.ai/...permissions
  61. ^1 ^2 ^3earendil-works. "Run Pi safely". Pi coding agent documentation, read 19 November 2026. github.com/...security.md
  62. ^Zed Industries. "Tool Permissions". Zed documentation, accessed 19 November 2026. zed.dev/...tool-permissions
  63. ^1 ^2The Linux Foundation. "Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md". Press release, 9 December 2025. linuxfoundation.org/...f-the-agentic-ai-foundation
  64. ^Kilo-Org. "kilocode-legacy" (archived repository and README). GitHub, read 19 November 2026. github.com/...kilocode-legacy
  65. ^1 ^2Brian Turcotte. "Thank you, Roo! We'll take it from here." Kilo Blog, 21 April 2026. blog.kilo.ai/...thank-you-roo
  66. ^1 ^2David DeSanto and Scott Breitenother. "AI on Your Own Terms: Anaconda Acquires Kilo Code". Anaconda blog, 15 July 2026. anaconda.com/...anaconda-acquires-kilo-code
  67. ^Earendil Inc. "Announcing Pi & Lefos". earendil.com, 8 April 2026. earendil.com/...announcing-pi-and-lefos
  68. ^Armin Ronacher. "RFC 0015: Pi Licensing". rfc.earendil.com, 30 March 2026. rfc.earendil.com/...0015
  69. ^OWASP Gen AI Security Project. "LLM01:2025 Prompt Injection". genai.owasp.org, accessed 19 November 2026. genai.owasp.org/...llm01-prompt-injection
  70. ^Yue Liu, Yanjie Zhao, Yunbo Lyu, Ting Zhang, Haoyu Wang, David Lo. "'Your AI, My Shell': Demystifying Prompt Injection Attacks on Agentic AI Coding Editors". arXiv:2509.22040, submitted 26 September 2025, revised 28 April 2026. arxiv.org/...2509.22040
  71. ^Cline. "Unauthorized npm publish of Cline CLI cline@2.3.0 with modified postinstall script to install openclaw" (GHSA-9ppg-jx86-fqw7). GitHub Security Advisory, 17 February 2026. github.com/...GHSA-9ppg-jx86-fqw7
  72. ^anomalyco. "Unauthenticated HTTP Server Allows Arbitrary Command Execution" (GHSA-vxw4-wv6m-9hhh, CVE-2026-22812). GitHub Security Advisory, 12 January 2026. github.com/...GHSA-vxw4-wv6m-9hhh
  73. ^OpenAI. "Sandbox bypass due to bug in path configuration logic" (GHSA-w5fx-fh39-j5rw, CVE-2025-59532). GitHub Security Advisory, 19 September 2025. github.com/...GHSA-w5fx-fh39-j5rw
  74. ^Zed Industries. "Symlink Escape in Agent File Tools" (GHSA-786m-x2vc-5235, CVE-2026-27967). GitHub Security Advisory, 25 February 2026. github.com/...GHSA-786m-x2vc-5235
  75. ^goose. "Arbitrary command execution in goose CLI via `goose review` via git core.fsmonitor" (GHSA-r5pp-p5r8-466r, CVE-2026-72718). GitHub Security Advisory, 24 July 2026. github.com/...GHSA-r5pp-p5r8-466r
  76. ^OpenHands. "Command Injection in Git Diff Handler" (GHSA-7h8w-hj9j-8rjw, CVE-2026-33718). GitHub Security Advisory, 23 March 2026. github.com/...GHSA-7h8w-hj9j-8rjw
  77. ^earendil-works. "Pi loads project-local extensions without approval" (GHSA-mqxh-6gq7-558m, CVE-2026-54325). GitHub Security Advisory, 8 June 2026. github.com/...GHSA-mqxh-6gq7-558m
  78. ^1 ^2Zed Industries. "External Agents". Zed documentation, accessed 19 November 2026. zed.dev/...external-agents
  79. ^Continue. "Continue has joined Cursor" (notice on the project's home page). continue.dev, accessed 19 November 2026. continue.dev
  80. ^1 ^2SWE-agent. "v2.0 Migration Guide". mini-SWE-agent documentation, accessed 19 November 2026. mini-swe-agent.com/...v2_migration
  81. ^1 ^2Kilo-Org. "cloud" LICENSE.md (Open Core Ventures Source Available License version 1.0). GitHub, read 19 November 2026. github.com/...LICENSE.md
  82. ^1 ^2Kilo-Org. "kilo-desktop" LICENSE.md (Open Core Ventures Source Available License version 1.0). GitHub, read 19 November 2026. github.com/...LICENSE.md
  83. ^openinterpreter. "openinterpreter" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...openinterpreter
  84. ^tailcallhq. "forgecode" (repository and LICENSE). GitHub, read 19 November 2026. github.com/...forgecode
  85. ^CodebuffAI. "freebuff" (repository, README and LICENSE). GitHub, read 19 November 2026. github.com/...freebuff
  86. ^OpenHands. "OpenAI-compatible gateway bypasses the Docker sandbox for new conversations, allowing agent commands to run on the host" (GHSA-h7m4-392w-vpjj). GitHub Security Advisory, 13 November 2026. github.com/...GHSA-h7m4-392w-vpjj
  87. ^TabbyML. "tabby" ee/LICENSE (The Tabby Enterprise license). GitHub, read 19 November 2026. github.com/...LICENSE
  88. ^anomalyco. "Figma skill materials" (NOTICE.md in the opencode repository). GitHub, read 19 November 2026. github.com/...NOTICE.md

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 6,739 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check before publication on November 19, 2026: 345 claims checked against 131 cited and independent sources; 9 material and 15 minor corrections applied before publication.

Cite this page: AI Wiki. "Best Open-Source Coding Agents." aiwiki.ai, updated 19 Nov 2026, fact-checked 19 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/best_open_source_coding_agents

Edit article