2.8 KiB

Raw Blame History

Warning

This repository is being monitored by SourceGraph developers, so I do not have any guarantees or promises about this not breaking right away. If they continue to change things around, I might stop updating sg-proxy altogether.

Note

If you are ratelimited and want to check the ratelimit for your account, go to the Cody chat and send a message. When over the limit, you should see an error like this:

Request failed: you exceeded the rate limit for chat_completions, only 3 requests are allowed per day at the moment to ensure the service stays functional. Current usage: 3. Retry after 2023-09-06 +0000 UTC

sg-proxy - free Claude

Requirements: Node, NPM.

Create a SourceGraph account (no phone required, just normal email/password).
Go to https://sourcegraph.com/, click on your profile icon in upper-right, then Settings, then go to Account->Access tokens, and generate a new token.
Clone/download https://gitgud.io/cncanon/sg-proxy/.
Open .env in the proxy folder and put your token there, so it looks like API_TOKEN="sgp_mytoken"
Run start.sh if you're on *nix, otherwise run start.bat. If you want to run manually, just npm install --no-audit and then node main.js.
Change completion to Claude in ST, set any API key, then set the proxy URL as http://localhost:3000/v1 and set any password.

If you want to use it remotely, you can use any tunneling service e.g. ngrok or cloudflared, for example just cloudflared tunnel --url http://localhost:3000 is enough to tunnel the proxy through Cloudflare, then you can access it via the given URL with the /v1 suffix.

Enjoy.

Updates

04/09/2023 - Fixed getting 403 on completion requests - seems like SourceGraph started checking for some additional headers, not a big deal.
30/08/2023 - Fixed two reported issues - broken streaming for multi-byte characters (non-English text) and misleading error for when there's an error. Now you should be able to stream non-English text (Japanese, Russian, etc) and, if sg-proxy catches an error, see the note about it in the message instead of ST never finishing the request.

Notes

Only tested with the model claude-2, older ones might work if their API allows to, maybe not.
The upper limit on the context size is unknown, please test. The initial value was at ~20K, but that was due to a (fixed) bug in the proxy. Some people have reported being able to use 50-80K context without issues.
The API has a daily rate-limit for requests. Some people have reported it as being at 100 requests, some - 50, but it might differ per-account. If you get a 429 error in console, just create a new account.

Contact: cncanon1@proton.me.