Agent Gateway
Use your GitHub Copilot models in the coding agents you already use: Claude
Code, Codex, Gemini CLI, Antigravity CLI, Kilo Code, Cline, DeepSeek Harness, Pi, Oh My Pi, Hermes, OpenClaw and
more. Agent Gateway
runs inside VS Code and gives them one local endpoint at
http://127.0.0.1:24444. You can also share your models with another machine,
and give agents web search and URL fetch.
Features
- One endpoint for every agent. Speaks the OpenAI (Responses and Chat
Completions), Anthropic Messages and Gemini APIs, so most agents work without
changes.
- One-click agent setup. Configure Coding Agents writes each agent's
settings for you, backs up every file first, and can undo its changes.
- Share across machines. Host the gateway through a dev tunnel, and add
other machines' gateways as remote upstreams.
- Web search and URL fetch. Agents that use built-in web search (Claude
Code, Codex, Gemini CLI) can search and read pages through Exa, Firecrawl,
Tavily, Jina AI or You.com, several engines without an API key.
- Easy troubleshooting. A status view, a 7-day error log, and a request ID
on every response and log line. Prompts and keys are never logged.
Getting started
- Install the extension. You need VS Code with GitHub Copilot Chat, signed in.
- When VS Code asks whether Agent Gateway may use the Copilot language
models, choose Allow. Without this, no models are served.
- Click Gateway in the status bar (or run Agent Gateway: Status) and
check that it says Ready.
- Run Agent Gateway: Configure Coding Agents, pick your agents and a
default model, review the changes and confirm.
The gateway runs while VS Code is open. To turn it off without uninstalling,
set agentGateway.enabled to false.
Supported agents
| Agent |
How it is configured |
| Claude Code |
Points Claude Code at the gateway; models are discovered automatically. Also sets CLAUDE_CODE_AUTO_MODE_SERVER=0, so auto mode uses Claude Code's own safety checks without a notice. |
| Codex |
Adds an agent-gateway provider and makes it the default. |
| Gemini CLI |
Set GOOGLE_GEMINI_BASE_URL=http://127.0.0.1:24444 (see manual setup). Google has moved personal accounts to Antigravity CLI; Gemini CLI still works here with an API key. |
Antigravity CLI (agy) |
Switches agy to its Gemini API key mode, then shows the two environment variables to set. agy offers only Gemini model names, so the gateway needs a matching model or a model alias. |
| Oh My Pi (omp) |
Adds the gateway as a provider with automatic model discovery; pick a model with /model. |
| Pi |
Adds the gateway with its model list; run Configure Coding Agents again after models change. |
| Hermes Agent |
Adds an agent-gateway provider; run hermes --provider agent-gateway or pick it in hermes model. |
| OpenClaw |
Adds the gateway with its model list. Checked against OpenClaw's documentation, not yet tested with a live install. |
| Kilo Code |
Adds the gateway with its model list; pick agent-gateway/<model> with /models. |
| Cline (CLI) |
Adds an OpenAI Compatible provider and makes it active; switch models with cline -m <model>. The Cline VS Code extension is set up by hand. |
DeepSeek Harness (dsh) |
Adds the gateway with its model list and makes your chosen model the default; removing it restores the previous default. |
| Cursor |
Shows manual steps. Cursor calls models from its own servers, so it needs a shared (dev tunnel) URL. |
| Any other agent |
Use the OpenAI, Anthropic or Gemini base URL; see manual setup. |
Each changed file is backed up next to itself as
<file>.agent-gateway-<timestamp>.bak. Configure Coding Agents >
Remove gateway reverts only what the gateway added.
Models for sub-agents and background tasks
Some agents send extra requests under their own model names, not the model you
picked. These are the names each agent sent in our tests:
| Agent (version tested) |
Model names sent besides your main model |
Change them with |
| Claude Code 2.1.295 |
claude-haiku-5.5 and claude-sonnet-5 (titles, summaries, helper and sub-agent requests) |
ANTHROPIC_DEFAULT_HAIKU_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, CLAUDE_CODE_SUBAGENT_MODEL in the env of ~/.claude/settings.json |
| Gemini CLI 0.32.1 |
gemini-3-flash-preview |
Built in; see step 3 below |
| Antigravity CLI 1.4.3 |
gemini-3.1-flash-lite-preview |
Built in; see step 3 below |
| Oh My Pi 18.10.3 |
claude-haiku-5.5 |
omp's model roles |
| Codex, Pi, Hermes, Kilo Code, Cline, DeepSeek Harness |
None: every request used your main model |
|
Everyone's Copilot model list is different. The gateway resolves each
requested name against your own list (Status, or GET /v1/models), in
this order:
- Your list has a model with exactly that name: it is used.
agentGateway.modelAliases maps the name, for example
{"claude-sonnet-5": "claude-sonnet-5.5"}: the target is used.- Gemini names that Gemini CLI and Antigravity CLI use internally
(
gemini-3-flash-preview, gemini-3.1-flash-lite-preview,
gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro,
gemini-3-pro-preview): the newest Gemini model in your list with "flash"
(or "pro") in its name.
- Any other name: the model in
agentGateway.fallbackModel, else
gpt-6-astra if your list has it, else the request fails. A Gemini request
with an unknown name fails instead. Run Set Fallback Model to choose it;
Status shows which fallback model is in use.
Every substitution is written to the Agent Gateway output panel, for
example Model fallback: requested "gpt-5.5" -> using "gpt-6-astra", and sent
back in the x-agent-gateway-model-fallback response header. Agents change
these names between versions, so check the output panel after updating one.
Your agent's own tools (commands, file edits, MCP servers) run on your
machine and work unchanged.
Some tools normally run on the model provider's servers. Copilot models in VS
Code have no such servers, so the gateway runs what it can itself:
| Server tool |
Through the gateway |
Web search (Anthropic web_search, OpenAI web_search, Gemini googleSearch) |
Run by the gateway through the engine you pick in Web Search and Fetch. Off until you pick one. |
Gemini URL context (urlContext, Gemini CLI's web_fetch) |
Run by the gateway, unless the same request also has the agent's own tools. |
Anthropic web_fetch |
Run by the gateway with the same engine. Only pages whose URL is already in the conversation are fetched. |
Code execution (Anthropic code_execution, OpenAI code interpreter, Gemini codeExecution) |
Not supported. Left out (Anthropic: the request fails). |
OpenAI file_search, image generation, remote mcp, hosted tool_search; Gemini Maps and File Search |
Not supported. Left out. |
| Other Anthropic server tools (MCP connector, tool search, advisor) |
Not supported. The request fails. |
A tool that is left out is never shown to the model, and the Agent Gateway
output panel notes it. None of the agents in the table above sends an
unsupported server tool in our tests.
Commands
All commands are under Agent Gateway in the Command Palette.
| Command |
What it does |
| Status |
Shows the gateway state, models, remotes and recent errors, with quick actions. |
| Restart Gateway |
Restarts the local gateway. |
| Change Port |
Moves the gateway to another port and offers to update your agents. |
| Set Fallback Model |
Chooses which of your models answers when an agent asks for a model you don't have. |
| Configure Coding Agents |
Sets up, updates or removes the gateway in your agents. |
| Web Search and Fetch |
Chooses the web search engine and its API key. |
| Share Gateway via Dev Tunnel |
Shares your gateway with another machine and manages the gateway API key. |
| Manage Remote Upstreams |
Adds, tests or removes another machine's gateway. |
Share models between machines
On the machine with Copilot, run Share Gateway via Dev Tunnel. It installs
and signs in to the devtunnel CLI if needed, offers to create a gateway API
key, and starts a tunnel whose URL stays the same across runs. Use Copy client
setup to send the URL (the key is never included).
On the other machine, run Manage Remote Upstreams > Add remote
upstream, paste the URL and enter the key. Its models appear as
<name>--<model>.
Anyone with the tunnel URL and key can use your Copilot models. Always set a
gateway API key before sharing.
Web search and URL fetch
Web search is off until you turn it on. Run Web Search and Fetch, pick an
engine, and enter its API key, or leave the key empty to use the engine's
free anonymous endpoint where there is one.
| Engine |
Without a key |
| Exa |
Search and fetch (rate-limited) |
| Firecrawl |
Search and fetch (rate-limited) |
| Jina AI |
Fetch only |
| Tavily |
Needs a key |
| You.com |
Needs a key |
Keys are stored in VS Code secret storage, never in settings.json. For Codex,
the command also offers to turn on its live search (web_search = "live").
Search queries, results and page text are never logged.
Settings
| Setting |
Default |
Description |
agentGateway.enabled |
true |
Run the gateway. |
agentGateway.port |
24444 |
Local port. Use Change Port to change it safely. |
agentGateway.logLevel |
info |
error, warn, info or debug. |
agentGateway.remoteUpstreams |
[] |
Other machines' gateways (managed by Manage Remote Upstreams). |
agentGateway.webSearch.provider |
none |
Web search engine (managed by Web Search and Fetch). |
agentGateway.webSearch.maxSearchesPerRequest |
5 |
Most searches per request (1-20). |
agentGateway.webSearch.model |
"" |
Model for Gemini CLI search requests when its model is unknown. |
agentGateway.modelAliases |
{} |
Map a requested model name to one of your models. |
agentGateway.fallbackModel |
"" |
Model for unknown model names (managed by Set Fallback Model). |
Troubleshooting
- Status shows Waiting for Copilot: make sure Copilot Chat is signed in and
that you allowed Agent Gateway to use the language models.
- Port in use: another program has port 24444. The gateway keeps retrying;
use Change Port to pick another one.
- A request failed: every response has a
request-id (and
x-request-id) header. Find that
ID in Status > Open error log or in the Agent Gateway output
panel. Set agentGateway.logLevel to debug for more detail.
- An agent stopped working after changing the port: run Configure Coding
Agents again.
Known issues
- Antigravity CLI fails with "invalid stream chunk: : ping": during a long
pause the gateway sends a standard keep-alive line on the stream, and
Antigravity CLI 1.4.3 does not accept it. Other agents are not affected.
Retry the request.
- Claude Code says auto mode "isn't eligible" for no-charge classifier
requests: auto mode's server-side checks need Anthropic's own API, so
through the gateway Claude Code runs its own checks, on your Copilot model.
Nothing breaks. Run Configure Coding Agents again to set
CLAUDE_CODE_AUTO_MODE_SERVER=0, which hides the notice.
Privacy and security
- The gateway listens on your computer only (loopback). Nothing is reachable
from other machines unless you share it through a dev tunnel.
- Prompts, request bodies, tool arguments, search queries and API keys are
never logged.
- API keys are kept in VS Code secret storage.
- Reference: API endpoints, manual client setup, how web
search and model fallback work, logs.
- Contributing: building, testing and the project's design.
- Changelog.
License
MIT
| |
