Self-hosted GitHub Copilot alternatives in 2026: maintained tools, team features and GPUs

Which Copilot alternatives run on your own server in 2026

A company that wants a self-hosted GitHub Copilot alternative has three kinds of maintained tool in October 2026. The open-source agents Cline (Apache 2.0) and Kilo Code (MIT) connect to a model server you run on premise; Tabby is an open-source server with an admin console; Tabnine Enterprise, Sourcegraph's Cody Enterprise and GitLab Duo Self-Hosted are commercial products installed in your own network. GitHub's own Copilot CLI can also work offline against a local model, so the code does not reach a cloud service.

Two tools that older guides recommend are no longer a choice for a new rollout. Roo Code's repository was archived by its owner on 15 May 2026; its README says the extension "was shut down on May 15th" and points to Zoo Code, a fork started by the Roo Code community, and to Cline. Continue's repository states that it "is no longer actively maintained and is read-only for all users" after a final 2.0.0 release. Our guide to a private coding assistant covers the models and their licences; this page compares the tools and what each needs from the server.

GitHub repositories, release pages and READMEs of each project; Tabnine, Sourcegraph and GitHub documentation; all read on 30 October 2026. Cline's README states that its JetBrains plugin is not open source. GitLab Duo from our GitLab article (GitLab docs 19.6, read 25 October 2026). Tabby's year from the order of its releases page.

One model or two for completion, chat and agents

Copilot does three jobs, and the alternatives cover different parts of them. Completion fills in code while a developer types. The suggestion has to arrive before the next keystroke, so it needs a small model trained for fill-in-the-middle and a tool that sends the code around the cursor. Our article on a code completion server sizes it. Tabby has a separate completion model slot, which for vLLM takes a vllm/completion endpoint and a prompt template for the model; Sourcegraph's model configuration lets an administrator assign a model to codeCompletion.

Chat is a developer asking about code: a few thousand tokens per question, answered at reading speed. Every tool in the table does it.

Agents are what most teams now mean by Copilot. Cline, Kilo Code, Zoo Code and Copilot CLI read files, run commands with the developer's approval and call the model again after each step, with prompts that grow to tens of thousands of tokens. GitHub requires tool calling and streaming from any model behind Copilot CLI and recommends at least 128k tokens of context. Kilo Code's Ollama page asks for "at least 32k to get decent results" and warns that qwen3-coder:30b "sometimes fails to call tools correctly"; Cline's Ollama page recommends its "Use Compact Prompt" setting for local models.

For the server this means two models when a team wants both completion and agents: a small completion model such as Qwen2.5-Coder-1.5B (3.09 GB, Apache 2.0) and an agent model such as Qwen3-Coder-30B-A3B or Qwen3-Coder-Next, each in its own engine instance. A team that only uses agents and chat needs one.

Team features such as admin console, sign-in and usage reports

With Cline, Kilo Code or Zoo Code, the settings live in each developer's IDE: a base URL, a key and a model name. Kilo Code's documentation states that its autocomplete requests are routed through the Kilo Gateway (read 31 October 2026), so that feature does not run on your server; where code must not leave the network, it stays off. Their READMEs describe no admin console for a self-hosted setup. Central control then comes from the server side. An LLM gateway in front of vLLM issues one key per developer, logs who called which model with how many tokens, and decides which models exist at all. Logs per developer are personal data under the GDPR; your data protection officer confirms their purpose and retention period before the rollout. Each extension's own telemetry is a separate question, which our article on telemetry in self-hosted AI tools answers for the server components.

Tabby brings these functions itself. Its news lists an admin UI (v0.9), team management and secured access (v0.7), team usage reports (v0.10), GitLab SSO and repository context from self-hosted GitHub and GitLab (v0.12), LDAP authentication announced for v0.24, indexing of GitLab merge requests as context (v0.30) and generic OAuth login (v0.32.0). The README calls it "Self-contained, with no need for a DBMS or cloud service". Its licence file puts everything outside the ee/ directory under Apache 2.0; the Tabby Enterprise licence for ee/ allows development and testing without a subscription, while production use needs "a valid Tabby Enterprise license for the correct number of user seats". The newest item in the project's news, of December 2025, concerns Pochi, a separate agent from the same team, and the releases page shows only nightly and alpha builds after v0.32.0.

Tabnine Enterprise is installed as "a Kubernetes cluster hosted on the customer's servers on their private network", and Tabnine states that private installations "can be deployed in a completely air-gapped environment". Tabnine writes that its team has no access to your servers but works with your team to set up and update the installation; the server specification depends on the number of users. Sourcegraph's Cody Enterprise, on a self-hosted Sourcegraph instance, takes your endpoint as an openaicompatible provider in its model configuration, and one model can carry both the autocomplete and chat capabilities. GitLab Duo Self-Hosted needs the Premium or Ultimate tier with the Duo Enterprise add-on; our GitLab Duo article covers its AI Gateway and models.

GitHub Copilot with a local model

GitHub documents three routes, and only one of them is described as local. Copilot CLI reads COPILOT_PROVIDER_BASE_URL and COPILOT_MODEL and supports Ollama, vLLM and other servers compatible with the OpenAI Chat Completions API. GitHub's authentication page states: "If you use your own LLM provider API keys (BYOK), GitHub authentication is not required." Without signing in, /delegate, the GitHub MCP server and GitHub code search are unavailable. With COPILOT_OFFLINE=true "No GitHub authentication is attempted", requests go only to your own provider and telemetry is switched off; GitHub adds that "Offline mode is only fully air-gapped if your BYOK provider is local or otherwise within the same isolated environment". GitHub's pages say the CLI "is available with all Copilot plans", and its README lists "An active Copilot subscription" as a prerequisite; the BYOK pages do not say whether that also applies when only your own model is used, so confirm this with GitHub before a rollout.

The Copilot app has supported bring-your-own-key since 23 June 2026, with LM Studio, Ollama and any OpenAI-compatible endpoint among the providers; on Business and Enterprise plans an administrator must enable Copilot CLI in the policy settings. For Copilot in the IDE, enterprise administrators can add custom models (public preview) for "Copilot Chat, Copilot CLI, and IDEs". That page does not say whether code completion can use them or whether requests pass through GitHub, so treat it as a cloud route until GitHub states otherwise. Claude Code, Codex and OpenCode with a local model are covered in our article on Claude Code with a local model, and Cursor in Cursor with a local model.

You can order the workstation or server alone, the setup of vLLM or Ollama with Cline or Kilo Code alone, or both together: the system delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Which setup fits which company

A studio of five developers needs one workstation and an agent on each laptop. Cline or Kilo Code points at the workstation's model server, and a reverse proxy with keys keeps the endpoint off the open network. Ollama is enough for a pilot if it runs version 0.35.0 or later, as our coding assistant guide explains; vLLM serves several developers at once more efficiently.

A department of 50 developers needs a server, vLLM with prefix caching and a gateway with one key per developer. If you plan such a server, ask us for a proposal with your developer count and Git platform. If the department wants an admin console and completion from the same product, Tabby is the open-source option. If it already runs GitLab Premium or Ultimate, GitLab Duo Self-Hosted keeps everything inside GitLab, at the price of the add-on licence. Our server for coding agents article sizes this range.

A regulated company that needs an air gap has four documented routes: Tabby, which needs no cloud service; Tabnine Enterprise in its air-gapped installation; GitLab Duo Self-Hosted with an offline licence; and Copilot CLI in offline mode with a local model. Models, drivers and containers then have to reach the server without internet access, as our guide to an air-gapped LLM server describes. Which rules apply when developers use such tools on company code is the subject of our guide to an AI coding assistant policy.

Which workstation or GPU server runs a Copilot alternative

The model sets the hardware, and the tool adds almost nothing. We count agent sessions by our users-per-card rule: 90 per cent of the memory the driver reports, less 3 GiB per card, less the weights, divided by the cache per session. That leaves 25.7 GiB on a 32 GB RTX PRO 4500, 61.5 GiB on the RTX PRO 5000 72 GB and 83.0 GiB on each 96 GB RTX PRO 6000. We count one generating session per developer at the busy hour as the upper bound.

Sessions of 64K tokens by our users-per-card rule, BF16 cache (FP8 cache in brackets; gpt-oss-20b with BF16 cache only). Weights from Hugging Face (gpt-oss-20b 13.76 GB, Qwen3-Coder-30B-A3B FP8 31.2 GB, Qwen3-Coder-Next FP8 80.4 GB), read 30 October 2026. Cache per 64K: 1.5 GiB (gpt-oss-20b, Coder-Next), 6 GiB (Coder-30B-A3B), by our arithmetic from each config.json. Speed column by the method of our coding-agents article, an upper bound by our estimate, computed for 1,792 GB/s per Workstation Edition or Max-Q and 1,597 GB/s per Server Edition (NVIDIA). TP=2: one copy of the model split across two cards (tensor parallelism).

The RTX PRO 4500 tier is memory-limited, and gpt-oss-20b, a 21B model with 3.6B active parameters under Apache 2.0, handles chat and short agent tasks; check its tool calls on your own repositories before you plan agent work on it. A completion model on the same card takes memory from the agent sessions: the 3.09 GB of Qwen2.5-Coder-1.5B on the RTX PRO 5000 72 GB leave 4 (9) sessions instead of 5 (10) by our arithmetic, before the second engine's own cache and overhead. Newer models need far less cache per session: by the same rule, this card holds Qwen3.6-35B-A3B (Apache 2.0) with ten sessions of 128K tokens, as our guide to Claude Code with a local model shows.

The mid tier fits a workstation or a rack server. Two RTX PRO 6000 Workstation Edition cards at 600 W each with flow-through coolers are the realistic maximum for that edition. Three or four RTX PRO 6000 Max-Q belong only in a chassis built for them with strong directed airflow: Puget Systems found in July 2025 that directly adjacent Max-Q cards have little room to draw air, and Exxact, in a page updated in March 2026, reports that four adjacent cards throttled without its custom airflow. In a rack for continuous operation we recommend the Server Edition; by our estimate, two cards with two AMD EPYC 9135 draw about 1.8 kW and four cards with two EPYC 9355 about 3.4 kW, so four cards need more than one 230 V 16 A feed, which carries about 2.9 kW at 80 per cent continuous load.

Lock-in and exit, with Roo Code as the example

Roo Code's shutdown shows where the lock-in sits. The shutdown on 15 May concerned the extension: the model server, the models and the gateway were not part of it, while the custom modes and prompts that teams had built in the extension stayed tied to it. Because the model server spoke a standard API, moving from Roo Code to another agent meant pointing that agent at the same server: a base URL, a key and a model name on each laptop.

Three habits keep that exit cheap. Keep coding rules and prompt files in the repository, where any tool can read them and they survive a change of tool. Serve models through vLLM or SGLang behind your own gateway, so that keys, logs and model choice do not depend on the extension. And for server products such as Tabby, Tabnine or Sourcegraph, check before the rollout how users, usage data and repository indexes can be exported, since those live in the product's own database.

We install the model server, the gateway and Cline or Kilo Code on your system; send us your developer count and the tools your teams use today, and we propose a configuration for your tasks and budget.

FAQ