Run a free DeepSeek web account as an OpenAI-compatible backend, then point opencode at it.

The DeepSeek web chat costs nothing. The metered platform API costs money. This project bridges the first one into the shape the second one has, keeps the whole thing on loopback, and paces every upstream request so an autonomous coding agent does not fire a turn every 200 ms and get the account flagged.

Two files do the work:

file	job
relay/extract.py	reads the login token out of a dedicated Firefox profile
relay/relay2.py	serves POST /v1/chat/completions on 127.0.0.1:20232

---

Contents

• How it works
• The tag contract
• Requirements
• Install
  • Step 1 — dedicated Firefox profile
  • Step 2 — extract the session
  • Step 3 — run the relay
  • Step 4 — wire up opencode
• Models
• Environment variables
• Limits
• Troubleshooting
• Credits
• License

---

How it works

opencode speaks OpenAI. The relay translates:

• in — the harness's system prompt and JSON tool schemas become a plain text tag contract (below), plus a short reminder every turn so long sessions do not drift into prose.
• upstream — the prompt is sent to chat.deepseek.com with the browser's own TLS/HTTP2 fingerprint, a solved proof-of-work header, and the chat session id from the previous turn, so the account sees one continuous conversation.
• out — the reply is parsed back into assistant text or tool_calls, and streamed to the harness as SSE deltas word by word.

Everything binds to loopback. No port forwarding, no public listener.

---

The tag contract

DeepSeek's web endpoint has no tool-calling API, so tools are described in text and read back out of text. Three reply dialects are tolerated, because the web client is not consistent about which one it produces:

<bash>{"command": "ls -la"}</bash>          plain tags
<DSML｜｜tool▁calls｜｜> …                    DeepSeek markup
< calls> …                                  reasoning model, bare

Unterminated tags are auto-closed before parsing, stray separators and broken closers are swept, and fenced code blocks are left alone — so a model that forgot to close a tag still yields a usable {"cmd": "ls -la"} instead of a turn of garbage.

---

Requirements

• Windows, macOS, or Linux
• Python 3.10+
• Firefox (real, installed — the extractor drives it)
• geckodriver on PATH
• A free chat.deepseek.com account, logged in once

---

Install

git clone https://github.com/Nbill27/opencode-deepseek-web.git
cd opencode-deepseek-web
pip install -r requirements.txt

requirements.txt pulls deepseek-pow (proof-of-work), wasmtime (its runtime), curl_cffi (browser TLS fingerprint), and selenium (token extraction).

Step 1 — dedicated Firefox profile

The relay needs a login token. It reads it from a Firefox profile you created for this, never your daily one.

1. Close every Firefox window.

2. Create a fresh profile and start it:

   firefox -profile ./ff-profile -no-remote

   Windows (default install path):

   & "C:\Program Files\Mozilla Firefox\firefox.exe" -profile .\ff-profile -no-remote

3. In that window, go to https://chat.deepseek.com and sign in.

4. Send one message so the session actually exists.

5. Close that Firefox completely. The token lives in memory until it is flushed to disk, and the extractor will read a stale one if the window is still open.

If your Firefox lives somewhere unusual, or geckodriver is not on PATH, tell the extractor:

export FIREFOX_BIN="/path/to/firefox"
export GECKODRIVER_PATH="/path/to/geckodriver"
export CHATBRAIN_FF_PROFILE="./ff-profile"

Step 2 — extract the session

python relay/extract.py

Writes relay/.websession.json (already in .gitignore — it is a live login, treat it like a password). Re-run this whenever the account gets logged out.

Step 3 — run the relay

python relay/relay2.py

It answers on http://127.0.0.1:20232. Check it:

curl http://127.0.0.1:20232/v1/models

{"object": "list", "data": [
  {"id": "deepseek-chat",     "object": "model"},
  {"id": "deepseek-reasoner", "object": "model"}
]}

A real completion:

curl http://127.0.0.1:20232/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-chat","stream":true,
       "messages":[{"role":"user","content":"say hi"}]}'

On Windows, start.ps1 does steps 2–3 and opens opencode for you:

./start.ps1

Step 4 — wire up opencode

~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "deepseek-web": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "DeepSeek (web)",
      "options": {
        "baseURL": "http://127.0.0.1:20232/v1",
        "apiKey": "unused-but-required"
      },
      "models": {
        "deepseek-chat": { "name": "DeepSeek Chat" },
        "deepseek-reasoner": { "name": "DeepSeek Reasoner" }
      }
    }
  },
  "model": "deepseek-web/deepseek-chat"
}

Then opencode and /model to switch.

---

Models

id	upstream	notes
deepseek-chat	V3 chat	fast, the default; no visible reasoning
deepseek-reasoner	R1	slower, thinking pass, emits the bare < calls> dialect

Both stream. usage in the response is estimated (token counts are approximated locally, not reported upstream) — good enough for cost display, not for billing.

---

Environment variables

All optional. Defaults are sane.

variable	default	meaning
CHATBRAIN_PORT	20232	listen port
CHATBRAIN_SESSION_FILE	relay/.websession.json	where the token lives
CHATBRAIN_LOG_FILE	relay/relay2.log	operational log
CHATBRAIN_REPLIES_LOG	relay/relay2-replies.log	raw conversations — delete before sharing
CHATBRAIN_MIN_GAP	2.5	seconds between upstream calls (log-normally jittered)
CHATBRAIN_JITTER	1.5	log-normal spread of that gap
CHATBRAIN_PAUSE_EVERY	15	calls before a long idle pause
CHATBRAIN_FAIL_COOLDOWN	30	seconds to back off after an upstream error
CHATBRAIN_IMPERSONATE	firefox147	curl_cffi TLS target
CHATBRAIN_UA	derived	override the User-Agent (escape hatch)
CHATBRAIN_FF_PROFILE	./ff-profile	extractor only
FIREFOX_BIN, GECKODRIVER_PATH	auto	extractor only

Do not set CHATBRAIN_MIN_GAP to 0 and then leave an agent running. That is exactly the pattern the pacing exists to prevent.

---

Limits

• This is a consumer chat account, not an API. Terms of service are on you. Do not publish the endpoint, share the box, or resell it.
• No image, file, or attachment input.
• No exact token counts — usage is estimated.
• Long conversations eventually hit the upstream context limit; the relay does not summarise for you.
• One account is one session chain. Two agents at once means interleaved context.
• Web upstream changes without notice. When replies come back as raw tags or empty text, check relay/relay2.log before assuming the relay broke.

---

Troubleshooting

symptom	cause
no session file at startup	run python relay/extract.py
extractor finds no token	Firefox still open, or wrong profile, or not logged in
every request 401 / "expired"	token died — re-extract; the relay warns once in the log
replies empty or full of raw tags	upstream format changed; look at relay2-replies.log
Connection refused on 20232	relay not running, or CHATBRAIN_PORT differs from baseURL
requests hang ~30 s	upstream rate limiting; raise CHATBRAIN_MIN_GAP
opencode shows no tools	model answered in prose; retry, and check the tag reminder in the log

---

Credits

Lineage: dsh-better-bridge by Samir Jamal — this relay is a fork of that codebase, hardened with a browser TLS transport, session chaining, log-normal pacing, and a more forgiving reply parser.

License

MIT — see LICENSE.