💡 The pitch

GitHub Copilot Chat is great, but you're locked to the models GitHub picks for you. This extension lets you bring your own Z.AI API key and chat with GLM-5.2 (1M context), GLM-4.7, GLM-5, plus free Flash and Vision models — right inside the Copilot Chat you already use. No Copilot Pro subscription, no second editor. There's also a @z-research agent that fetches dozens of cited sources and writes you a report.

flowchart LR
    A["You, in VS Code"] -->|"type a prompt"| B[Copilot Chat]
    B -->|"model picker"| C["Z.AI GLM-5.2 / 4.7 / Flash / Vision"]
    C -->|"your API key"| D["https://api.z.ai"]
    D -->|"streamed reply"| B

🔥 Why bother

You already love Copilot Chat. Now imagine it powered by Z.AI's GLM models — a 1M-context flagship, a free Flash tier, vision models that read screenshots, and a built-in deep-research agent that hands you a cited report.

📊 GitHub Copilot vs This Extension

Not a replacement — this extension adds models into Copilot Chat. Think of it as unlocking the model picker.

You pay Z.AI per-token (or use the free Flash models). No middleman subscription.

What Is This?

Z.AI for GitHub Copilot Chat is a VS Code extension that registers Z.AI GLM models (including GLM-5.2, GLM-5.1, GLM-5, and GLM-4.7) into GitHub Copilot Chat through the official VS Code Language Model Chat Provider API.

You pick a Z.AI GLM model from the Copilot Chat model picker the same way you would pick GPT-4 or Claude. Enter your Z.AI API key once, and that's it. No Copilot Pro or Enterprise subscription needed.

✨ Features

• Bring your own key. Enter your Z.AI API key once, every model is unlocked.
• Live model list. The extension pulls the current Z.AI lineup on startup, so new models appear automatically.
• Works offline too. If the API is unreachable, a bundled table with accurate per‑model token limits takes over.
• Per‑model token limits. Context window and max output tokens are set per model, not as one blunt global cap.
• Tool calling. Tool schemas are forwarded using OpenAI‑compatible chat completions, so agents keep working.
• Reasoning debug. Opt‑in reasoning_content logging to the Z.AI output channel, for when you want to see the model think out loud.
• One‑click diagnostics. A markdown report showing exactly which models VS Code has registered.
• Deep research agent. The @z-research chat participant runs Z.AI's MCP Web Search and Web Reader across several iterations to produce a cited research report. See Deep Research below.
• Progress you can actually see. Each completed search query is pushed to the chat as a progress update, not one big batch at the end.
• Built to survive Z.AI's quirks. The extension handles double‑encoded JSON responses, retries on rate‑limit (429) with exponential backoff, and enforces a per‑call timeout so a single hung MCP call can't freeze your run.
• Junk URL filter. Instagram, TikTok, YouTube, asset CDNs, and "how to host" guides are dropped at the candidate stage. That alone saves a 30s timeout per junk URL.

🔬 Deep Research

The extension registers Z.AI's remote MCP servers for Web Search and Web Reader and exposes them to Copilot Agent. The @z-research participant then runs them across several iterations to produce a multi‑source, cited research report. The result goes well past the two or three links the built‑in Copilot web search returns.

The pitch in one line: type @z-research who is winning the on‑device LLM race in 2026, wait a few minutes, get back a cited markdown report with dozens of ranked sources.

How it works

Z.AI's Web Search and Web Reader are MCP servers, not REST endpoints. Usage is billed against the GLM Coding Plan's shared monthly MCP quota, not the general API balance, so no top-up is needed.

@z-research participant (multi-iteration, hundreds of sources)

For thorough research the participant runs its own loop, which lets it bypass Copilot's per-turn tool-call cap:

1. Plan: the synthesis model generates 5 to 10 search-engine-friendly queries. These favour concrete entities, action-oriented phrasing, and a 3 to 8 keyword sweet spot.
2. Filter: candidates pass through the junk URL filter (Instagram, TikTok, YouTube, asset CDNs, "how to host" guides) and are deduplicated by URL across queries.
3. Search: queries run in parallel, bounded by zai.research.concurrency, via the webSearchPrime MCP tool. Each call has a 30s timeout, and 429 responses retry with exponential backoff.
4. Read: top URLs are fetched through the webReader MCP tool with two-tier caching. Junk URLs are skipped before the read.
5. Rank: sources are deduped and scored using a BM25-style term overlap with a recency boost. Only the top 25 most relevant sources go to synthesis.
6. Expand: if budget remains and coverage is thin, a gap analysis of the previous round's top results drives 3 to 5 follow-up queries. There are no random re-rolls.
7. Synthesise: the top 25 sources are chunked (16K chars each), each chunk is summarised (map), and a final cited report is produced (reduce). The synthesis LLM is steered to maximise what the sources actually cover, rather than giving up if the user's angle isn't a perfect match.

Progress is pushed to the chat as each search query completes, not as one big update at the end.

flowchart TD
    U["@z-research topic"] --> Plan[Plan: 5-10 quality queries]
    Plan --> Filter[Filter: dedup + junk URL]
    Filter --> Search[Parallel search via MCP, 30s timeout each]
    Search -->|429| Retry[sleep 1s/2s/4s + retry]
    Search --> Read[Parallel webRead via MCP + cache]
    Read --> Rank[Dedupe + BM25 rank top 25]
    Rank --> Done{Budget ok?}
    Done -- no --> Synth[Map-reduce synthesize 16K chunks]
    Done -- yes --> Expand[Expand: gap analysis, 3-5 new queries]
    Expand --> Search
    Synth --> Out[Render report + clickable sources]

Usage:

@z-research <topic> is the single entry point. There are no slash commands.

• Default mode (quick): about 20 sources, 1 to 2 iterations, roughly 3 to 4 minutes end-to-end.
• Deep mode: include keywords like deep, thorough, comprehensive, lengkap, or menyeluruh in your prompt to opt in. Up to 100+ sources and 5 iterations.

Examples:

• @z-research pricing kompetitor SaaS WhatsApp di Indonesia 2026 (quick mode)
• @z-research deep research complete state of agentic coding tools June 2026 (deep mode)

Typical performance (measured on real Coding Plan runs):

• 8 to 15 search queries
• 30 to 130 candidate URLs after the junk filter
• about 25 sources actually read and ranked
• 3 to 5 LLM synthesis calls (1 reduce plus N chunk summaries)
• 2 to 4 minutes wall-clock per run

First-time setup

No MCP setup required. The @z-research participant calls the Z.AI MCP HTTP endpoints directly via fetch(). The tools are not registered with VS Code's MCP infrastructure, so they are invisible to Copilot Agent and other chat participants. That matters: it stops Copilot Agent from auto-discovering and invoking the tools during a regular chat, which is what caused stuck sessions in earlier versions.

The only prerequisite:

1. Open the Command Palette and run Z.AI: Set API Key (if you haven't already).
2. Type @z-research <topic> in Copilot Chat.

The participant will display a clear error if the API key is not set.

The final response is a markdown report with inline [n] citations and a clickable Sources list.

📚 Implementation history: doc/deep-research-journey.md has the full build log, covering 9 phases, 28 production bugs with root-cause analysis, 18 lessons learned, and the final architecture.

Requirements

• VS Code 1.120.0 or higher with the Language Model Chat Provider API
• GitHub Copilot Chat extension. Install it from the marketplace. This extension only adds models into Copilot Chat.
• Sign in to GitHub Copilot Chat. A personal GitHub account is enough; no Copilot Pro or Enterprise needed for BYOK.
• A Z.AI API key. Get one at z.ai. The free Flash models mean you can start without paying anything.

⚡ Quick Start

Five minutes from zero to your first GLM reply.

1. Install GitHub Copilot Chat from the marketplace if you haven't already.
2. Install this extension (or press F5 in the repo to launch an Extension Development Host).
3. Open GitHub Copilot Chat (click the Copilot icon in the sidebar or press Cmd+Shift+I / Ctrl+Shift+I).
4. Click the model picker (current model name) → Manage Models…
5. Select Z.AI.
6. Press Enter to accept the default Group Name.
7. Enter your Z.AI API Key when prompted. VS Code stores it securely as a secret.
8. Choose the models you want available.
9. Select any Z.AI model from the picker and start chatting.

💡 Tips:

• Registered models show up automatically in the Copilot Chat model picker.
• If a model appears in the Language Models view but not in the chat picker, hover its row and click the eye icon (👁) to enable visibility.

Commands

Once installed, Z.AI models appear directly in the GitHub Copilot Chat model picker with no extra commands. The easiest way to manage your API key is Settings, Language Models (gear icon ⚙).

For advanced usage, you can also run these commands via the Command Palette (Cmd+Shift+P / Ctrl+Shift+P):

Note: The native BYOK flow via Language Models (gear icon ⚙) is recommended.

Coding Plan quota

When your API key belongs to a Z.AI Coding Plan subscription, the extension shows a quota indicator $(graph) Z · NN% on the right side of the status bar:

• Hover the indicator to see a graphical SVG donut chart with two concentric rings. The outer ring is the weekly quota, the inner ring is the rolling 5-hour quota. Each ring is colour-coded: blue (normal), yellow at 80% or above, red at 95% or above. Below the chart: usage percentages and reset countdowns.
• Click the indicator to toggle the status-bar text between the 5-hour and weekly view.
• The indicator background turns yellow at 80% usage and red at 95%.
• Z.AI: Manage Provider → Show Quota opens a detailed markdown report with all quota windows.

The quota is fetched from https://api.z.ai/api/monitor/usage/quota/limit and auto-refreshes every 5 minutes (configurable via zai.quotaRefreshInterval).

If quota data is unavailable (for example, no API key set, or the key doesn't belong to a Coding Plan), the status bar shows a persistent $(graph) Z.AI quota item with a tooltip linking to Z.AI: Set API Key.

Settings

Troubleshooting

"No utility model is configured for 'copilot-utility-small'"

This is a VS Code 1.128 behavior change. When a BYOK model is set as the main agent, VS Code no longer falls back to a Copilot-provided utility model for background tasks (chat titles, commit messages, intent detection).

The extension fixes this automatically. On VS Code 1.128+, if no utility model is configured, the extension sets chat.byokUtilityModelDefault = "mainAgent" in your global settings on first activation and shows a brief toast to confirm. No manual action needed.

If you still see the error (for example, after resetting settings), reload the VS Code window (Cmd+Shift+P → Reload Window) to let the extension re-apply the fix.

If you want to configure a different model for utility tasks, set chat.utilitySmallModel or chat.utilityModel manually — the extension won't overwrite an explicit configuration.

Full patch notes: doc/vscode-128-byok-utility-model.md — root cause analysis, failed attempts, verified enum values, and code walkthrough.

"Request timed out for glm-5.1" / "Connection timed out after …"

Flagship 200K-context models (glm-5.1, glm-5, glm-5-turbo, glm-4.7) have noticeably higher cold-start latency than the smaller models. On long or busy sessions they can take 60 to 120s to send the first token.

The extension already mitigates this automatically:

• zai.requestTimeout defaults to 180000ms (3 min). It was 120000ms in 0.1.x.
• The effective connection timeout is auto-scaled to 1.5× for 200K flagship models, so 180s base becomes 270s.
• The inactivity timer scales the same way, with a 90s minimum floor (was 30s).

If you still hit timeouts:

1. Retry. Z.AI servers sometimes spike under load; the same prompt may succeed a few seconds later.
2. Increase zai.requestTimeout in Settings (for example, 300000 = 5 min max).
3. Try a faster model like glm-4.5-flash or glm-4.7-flash for code-completion or quick-edit tasks.
4. Clear chat history to reduce input token count. Large prefill is the main driver of cold-start latency.
5. Check the Z.AI Output channel. Every request logs [Timeout config: model=X flagship=Y multiplier=Z× connectionTimeout=…], so you can confirm which budget was applied.

If the issue persists with zai.requestTimeout = 300000 and a small context, the Z.AI API itself is the bottleneck. Try a different Z.AI region or plan, or contact Z.AI support.

"Z.AI model not selectable in the model picker" / "Can't pin Z.AI model"

The Z.AI extension only sends the official LanguageModelChatInformation fields to VS Code. Non-API fields like category and isUserSelectable are not part of the public VS Code API, and sending them can make the picker misbehave or crash. See doc/vscode-126-chatmodel-picker-crash.md for the original incident.

If the model picker doesn't show your Z.AI models or they can't be pinned:

1. Make sure Z.AI models are enabled in the picker. Open the picker, search for "Z.AI", and click the eye (👁) icon to enable visibility. The eye icon is in the Language Models view (gear icon ⚙ → Z.AI) and toggles whether the model is listed in the picker.
2. Pin a model as default. Set zai.defaultModel in your user settings (e.g. glm-5.2). The extension marks that model as isDefault: true so VS Code highlights it in the picker and seeds new chat sessions with it.
3. Reload the window after changing zai.defaultModel (the model list is cached per-window).

If the picker still misbehaves:

• Open Developer Tools (Cmd+Shift+P → "Developer: Toggle Developer Tools") and look for console errors when you open the picker.
• Check Output → Z.AI for any error logs from the model provider.
• File an issue with the console error and your VS Code version.

"@z-research: Z.AI API key is not set"

The @z-research participant calls the Z.AI MCP HTTP endpoints directly. It needs your Z.AI API key to authenticate.

1. Run Z.AI: Set API Key from the Command Palette.
2. Re-run @z-research <topic>.

Note: In v0.3.0, a Z.AI: Setup MCP Servers command was used to write mcp.json. This command was removed in v0.3.1: the extension now calls the Z.AI MCP endpoints directly via HTTP. If you have zai-web-search-prime or zai-web-reader entries in your mcp.json, you can safely remove them; they are no longer needed.

"@z-research: No usable sources were found"

Several possible causes:

1. Search queries returned empty results. The topic may be too niche. Try rephrasing with concrete keywords.
2. All top URLs were filtered as junk. If every returned URL was social-media or an asset CDN, the filter dropped them all. Try a more specific topic with named entities, for example "World Archery 2024 registration rules" instead of "archery registration".
3. Every read timed out. Check the Output channel for [mcp-tools] Timeout (30000ms) entries. If there are many, your Z.AI MCP servers may be rate-limited or unreachable.

The diagnostic log is in Output → Z.AI Research and includes every search query, every read, and the parsed result count.

"@z-research takes 4+ minutes"

That's the normal end-to-end time for a deep-mode run. The wall-clock time is bounded by:

• Search phase: bounded by zai.research.concurrency (default 3) and the 30s per-call timeout.
• Read phase: up to about 25 source reads in parallel, again 30s timeout each.
• Synth phase: 3 to 5 LLM calls (1 reduce plus N chunk summaries) on the synthesis model. With glm-5.2 (1M context), this is fast.

To shorten: use quick mode (omit deep / thorough / menyeluruh keywords from your prompt), lower zai.research.maxSources, or pick a smaller synthesis model.

"@z-research hits MCP rate limit (-429)"

The participant automatically retries with exponential backoff (1s / 2s / 4s) on rate-limit responses. If you see persistent 429s in the log:

• Lower zai.research.concurrency to 2 (default is 3).
• The Z.AI Coding Plan has a monthly MCP quota (Lite=100, Pro=1K, Max=4K calls). Check your usage in the Z.AI dashboard.

"@z-research search query stuck/hangs"

Each MCP call has a 30s per-call timeout. A hung call is logged as [mcp-tools] Timeout (30000ms) for search "..." giving up on this query and the orchestrator continues. If you see many timeouts:

• Z.AI's MCP server may be having a transient issue. Try again in a minute.
• The query itself may be problematic. If a particular query consistently times out, consider rewording it in your prompt.

Models

The extension fetches the live model list from:

https://api.z.ai/api/coding/paas/v4/models

Because the Z.AI API returns model IDs only, a bundled metadata table provides context window and max output tokens per model. If the live fetch fails, the bundled list is used as a fallback.

VS Code and Copilot read separate input and output metadata fields for the UI. GLM models can have very large output limits, so the extension advertises a small response reserve. That keeps the Language Models table, the model picker tooltip, and the chat context indicator consistent, while still sending each model's full bundled max output limit to the Z.AI API.

Bundled model limits

Set zai.maxInputTokens or zai.maxTokens to a non-zero value to override the bundled defaults globally.

All models use the OpenAI-compatible chat completions endpoint:

https://api.z.ai/api/coding/paas/v4/chat/completions

Development

# Install dependencies
npm install

# Compile TypeScript
npm run compile

# Watch mode
npm run watch

Press F5 in VS Code to launch an Extension Development Host with the extension loaded.

To package a .vsix for local install:

npm run package

Tests

Two test suites, both using Node's built-in test runner:

# Quota module
npx tsx --test src/test/quota.test.ts

# Deep research modules
npm test

The npm test runner covers 9 vscode-free pure modules: mcpInputBuilders (5), mcpResponseParser (15), mcpRateLimit (6), mcpTimeout (5), mcpToolNameResolver (9), junkUrlFilter (10), ranker (5), budget (5), cache (5), plus URL dedup (5). 75 tests, 100% pass rate, runs in ~600ms.

❓ FAQ

Do I need Copilot Pro?

No. A free GitHub account is enough. Copilot Chat itself is free to use, and this extension adds Z.AI models into it via the Language Model Chat Provider API. You only pay for Z.AI API usage (or use the free Flash models).

Is this an official Z.AI or GitHub extension?

No. This is an independent, open-source project. Z.AI and GitHub Copilot are not affiliated with this extension. It uses their public APIs.

Do the models work with Copilot Agent / tool calling?

Yes. Tool schemas are forwarded using OpenAI-compatible chat completions, so agents and tool calling keep working.

Is it really free?

The extension is free and open source (MIT). Z.AI offers glm-4.5-flash (text) and glm-4.6v-flash (vision) for $0. Other models are pay-per-token. Check z.ai for current pricing.

What's the difference between this and Copilot's built-in web search?

Copilot's built-in search returns 2–3 links. The @z-research participant runs a multi-iteration loop that fetches 20–100+ sources, reads them, ranks by BM25 + recency, and writes a cited markdown report. It's a different class of research.

Can I use this without Copilot Chat?

No. This extension only adds models into GitHub Copilot Chat. You need the Copilot Chat extension installed and signed in.

💬 Community

• Found a bug or have a feature idea? Open an issue
• Want to contribute? Pull requests welcome — see Contributing below
• Star the repo ⭐ if this saved you a subscription or a research afternoon. Word of mouth is how side projects reach the people who need them.
• Leave a review on the VS Code Marketplace — it helps others find the extension.

🪶 Also from this publisher

Looking for OpenCode Zen / Go models (DeepSeek, Kimi, Qwen, MiMo, MiniMax) in Copilot Chat? Check out OpenCode Copilot Chat — 5,000+ installs, 30+ models.

Contributing

Issues and pull requests are welcome. If you are planning a bigger change, please open an issue first so we can talk through the approach.

If this extension saved you a Copilot Pro subscription, a coffee, or a research afternoon, the nicest thing you can do is star the repo and tell one friend who would care. Word of mouth is how side projects like this reach the people who need them.

Contributors

• @nik13513513 (Alex Kor): Z.AI Coding Plan quota tracking, graphical SVG donut tooltip, quota auth error handling, and the test suite ([PR #2](https://github.com/ltmoerdani/zai-copilot-chat/pull/2), [PR #3](https://github.com/ltmoerdani/zai-copilot-chat/pull/3))

License

MIT. See LICENSE for details. Build things, ship things, share things.