Repository navigation
Releases
v0.2.6
What's Changed in v0.2.6
- Double-encoded
Bash.commandwith unescaped inner double quotes: Moved double-encoded{"command": "..."}unwrapping in_fit_argumentsto run afterwrap_key/ARG_ALIASES, with a regex fallback when unescaped inner shell double quotes break JSON/lenient parsing (fixes residualExit code 127: command not found: command:errors). - Control-character restoration in
Bash.command: Restores JSON-decoded\x08/\x0c(from unescaped\b/\fin shell regexes) back to\b/\fand strips stray C0 control characters (preserving\n,\r,\t) so Claude Code's Zod validator no longer rejects regex commands.
Install / Upgrade:
pipx upgrade venice-bridge # or: pip install -U venice-bridge==0.2.6v0.2.5
What's Changed in v0.2.5
- Explicit
webEnabled: false/webScrapeEnabled: false/xSearchEnabled: falseby default: Eliminates Venice's automatic web-search hijacking on coding & tool-calling turns unless:webor:xsearchis requested. - Claude Code
2.1.329tool-schema parity: Added Claude Code's snake_case argument names (file_path,old_string,new_string,replace_all,run_in_background,skill) toARG_ALIASESand"agent"toSAME_TOOL["task"]. - New
_route_callaliases:SQL_ALIASES(sql_query,sqlite_query,db_query,execute_sql,run_sql,sql, etc.) → read-onlysqlite3 -readonly <db> <sql>viaBashDATE_ALIASES(get_current_date,get_time, etc.) →date -u '+%Y-%m-%d %H:%M:%S UTC'viaBash- Deferred
grep/glob→ boundedgrep -rn ... | head -n 50andfind ... | head -n 50viaBash
- Double-encoded
commandunwrap &TOOL_REMINDERcleanup: Unwraps{"command": '{"command": "..."}'}payloads and removes the literal"example_tool"placeholder fromTOOL_REMINDER. claude-venice--resumebypassPermissionsfix: Pairs--dangerously-skip-permissionswith--permission-mode bypassPermissionsby default so resumed sessions stay autonomous (optionalcvshort alias documented).
Install / Upgrade:
pipx upgrade venice-bridge # or: pip install -U venice-bridge==0.2.5v0.2.4
Added
- Proactive Clerk JWT refresh. The session token is refreshed in the background when fewer than
VENICE_BRIDGE_PROACTIVE_REFRESH_Sseconds (default 30) remain, instead of on the request path
(7.5% of measured upstream opens had waited ≥ 0.5 s for auth). GET /stats(JSON, or?format=html): live request, model, TTFT, retry, cache, auth, pool and endpoint counters. Counts only; no prompts, cookies or tokens.tool_choice(auto,none,required, named) andparallel_tool_calls: falseon
/v1/chat/completions; Anthropictool_choiceanddisable_parallel_tool_useand Responses
parallel_tool_callsare forwarded by the adapters.response_format(json_object,json_schema) and Responsestext.format: the model is told to answer in JSON and fenced or padded JSON is cleaned up.stop/stop_sequences(no-tools path): output stops before the first match, safe across stream chunks; the Anthropic adapter reportsstop_reason: "stop_sequence".- Anthropic thinking blocks: reasoning is streamed as
thinkingcontent blocks on/v1/messages. - Opt-in multi-account cookie pool (
VENICE_BRIDGE_COOKIE_POOL, a 0600 file outside the repo, one cookie per line; seecookie-pool.example.txt). When the active account hits its quota (HTTP 429) the bridge moves to the next account until the reset. Unset = single-account behavior, unchanged. Accounts appear only asacct#Nin logs and/stats. - Multimodal image & document attachments across OpenAI
/v1/chat/completions(image_url,file), Anthropic/v1/messages(image,documentbase64), and OpenAI/v1/responses(input_image,input_file). Requests carrying attachments are sent asmultipart/form-datato/api/inference/chat/agents/ad-hoc; text-only requests stay unchanged on the plain chat route. Enforces a 17-type MIME allowlist (png/jpg/gif/webp + txt/md/csv/rtf/json/yaml/xml/pdf/docx/xls/xlsx), rejects remotehttp(s)://URLs with HTTP 400 (remote URLs not supported, send base64), enforces per-file (10 MB) and total (20 MB) size limits with HTTP 413 (VENICE_BRIDGE_MAX_ATTACHMENT_BYTES,VENICE_BRIDGE_MAX_TOTAL_ATTACHMENT_BYTES), skips response cache for multimodal turns, and never logs raw bytes/base64.- Known limitation: Because attachments require
/api/inference/chat/agents/ad-hoc, combining an image attachment with client tool calls in the same turn can collide with Venice's own server tools (measured:2/5bashtool calls survived ondeepseek-v4-1-flash-uncensoredwith an attached PNG; text-only tool turns are unaffected). Only attachments sent after the last assistant turn are uploaded; older ones become a placeholder, so the rest of the conversation returns to the plain chat route.
- Known limitation: Because attachments require
- OpenAI-compatible Speech-to-Text (
POST /v1/audio/transcriptions) backed by Venice/api/inference/asr(nvidia/parakeet-tdt-0.6b-v3default;json,text, andverbose_jsonformats; mp3/wav/m4a/webm/ogg/flac allowlist; 25 MB cap viaVENICE_BRIDGE_MAX_AUDIO_BYTES→ HTTP 413; 0 chat/image quota spent). - OpenAI-compatible Image Generation (
POST /v1/images/generations) backed by Venice in-chatveniceai:image_generation.image_chunk(n=1,b64_jsondefault orurlasdata:URI, no file hosting; decrements bothchatandimagequota by 1). Both endpoints add count-only counters underGET /stats(endpoints.audio_transcriptions,endpoints.image_generations) and never log or cache raw audio/image bytes.
Install: pipx install venice-bridge==0.2.4
v0.2.3
Added
claude-venice --leanlauncher flag: disables all MCP servers via--strict-mcp-config --mcp-config '{"mcpServers":{}}'to keep the[TOOLS]block compact when a project carries.mcp.json(default behavior unchanged).
Fixed
- Runaway repetition loops cut early. In tool mode, replies exceeding 40 tool-call envelopes or 150,000
characters are cut at the 41st envelope / cap and parsed as complete instead of hanging for minutes
(VENICE_BRIDGE_LOOP_MAX_CALLS,VENICE_BRIDGE_LOOP_MAX_CHARS). - Code written instead of run. A reply with no tool call that holds a fenced code block while a shell tool
is declared now goes to the judge, which can ask the model to run the command instead of showing it. bash_toolis routed to the client's shell tool.
Install: pipx install venice-bridge==0.2.3
v0.2.2
A small fix release for tool calling, plus a changelog.
Fixed
- Bare-string tool calls are no longer dropped. When a model answers a tool call with a bare string, it reaches the bridge as
{"content": "..."}. If the tool has nocontentparameter and exactly one required string parameter, the string now goes into that parameter instead of being dropped (dropped args not in schema: ['content']). 7 new self-check cases.
Docs
- CHANGELOG.md and a What's new section on the documentation site.
- The argument-fitting docs now cover envelope unwrapping and bare-string content.
Install
pipx install venice-bridge # or: pip install venice-bridge| File | For |
|---|---|
venice_bridge-0.2.2-py3-none-any.whl |
pip install this file directly: gives venice-bridge, venice-adapters and claude-venice. |
venice_bridge-0.2.2.tar.gz |
Source distribution (same files as the wheel, buildable with pip). |
venice-bridge-0.2.2.zip |
The full repository at this tag (docs, examples, tests). Unzip, pip install curl_cffi, follow the README Quick start. |
SHA256SUMS.txt |
Checksums of the three files above. |
Upgrading from 0.2.1: replace venice_bridge.py (or pipx upgrade venice-bridge) and restart the bridge. No configuration changes.
Unofficial and not affiliated with Venice.ai. It automates your own logged-in web session; see the README warning.
v0.2.1
A small fix release for tool calling on the Venice web models.
Fixes
- Double-nested tool arguments are unwrapped. Models often wrap the real arguments in an extra envelope:
{"arguments": {...}},{"name": "...", "arguments": {...}}, a JSON string inarguments, or Anthropic-style{"input": {...}}/{"parameters": {...}}. The bridge now unwraps up to three levels before checking the tool's schema (only when the schema has no field of that name), so the inner command is no longer dropped asdropped args not in schema: ['arguments']. - Format reminder: the model is told to put parameters directly in
argumentsand to use a heredoc for multi-line SQL or Python, which avoids broken JSON quoting in long shell commands. - 5 new self-check cases (
nested_arguments_unwrap_checks);docs/TOOL_ROUTING.mddocuments the unwrap.
Install
| File | For |
|---|---|
venice-bridge-0.2.1.zip |
The source of this release (same as the tag). Unzip, pip install curl_cffi, then follow the README Quick start. |
SHA256SUMS.txt |
Checksum of the zip. |
Upgrading from 0.2.0: replace venice_bridge.py and restart the bridge. No configuration changes.
Unofficial and not affiliated with Venice.ai. It automates your own logged-in web session; see the README warning.
v0.2.0 — Claude Code and Codex, local models, venice-auto
venice-bridge is a free, open-source local bridge for macOS that lets OpenCode, Claude Code, Codex CLI and other OpenAI- or Anthropic-compatible coding agents use a venice.ai web subscription, with working tool calls.
Highlights
- Claude Code and Codex CLI work through a new adapter,
venice_adapters.py(127.0.0.1:18892), which translates Anthropic Messages and OpenAI Responses into Chat Completions for the bridge. Standard library only. claude-venicelauncher: runs Claude Code against the adapter in an isolated profile (~/.claude-venice), so your normal Claude Code install and login are untouched. It skips permission prompts by default; read the warning in the README before using it.- Local models by prefix:
ollama/<model>,lmstudio/<model>and any OpenAI-compatible server frombackends.jsonare served by the same bridge and listed in/v1/models. venice-auto: a virtual model that uses Venice normally and a local model while the daily web quota is used up.- Opt-in Venice features: web search (
:web,:xsearch) and reasoning on no-tool requests (:think, returned asreasoning_content). temperatureandtop_pare forwarded; credit-only and end-to-end-encrypted models are hidden from/v1/models; byte-identical repeat requests are answered from an in-memory cache without spending quota.
Fixes
- A local model asked through the adapter is never sent to Venice when its backend is down (it fails with 502).
reasoning_effortno longer turns reasoning on for requests that declare tools (avoids an HTTP 400 in agent loops).venice-bridge/venice-autoresolves likevenice-auto.
Install
| File | For |
|---|---|
venice-bridge-0.2.0.zip |
The source of this release (same as the tag). Unzip, pip install curl_cffi, then follow the README Quick start. |
SHA256SUMS.txt |
Checksum of the zip. |
You can also git clone https://github.com/nsozturk/venice-bridge.git. No installer is needed: it is two Python scripts and a bash launcher.
Upgrade notes
- New optional process: run
python3 venice_adapters.pynext to the bridge if you use Claude Code or Codex. - New environment variables:
VENICE_WEB,VENICE_X_SEARCH,VENICE_REASONING,VENICE_EXCLUDE_MODELS,VENICE_AUTO_MODELS,VENICE_BRIDGE_BACKENDS,VENICE_BRIDGE_CACHE*,VENICE_BRIDGE_HOST, and the adapter'sVENICE_ADAPTER_*. All have safe defaults; nothing changes unless you set them. /v1/modelsnow listsvenice-autofirst and may list local models; credit-only models are no longer listed (they can still be requested by id).
Unofficial and not affiliated with Venice.ai. It automates your own logged-in web session; see the README warning.