Synapse
The open-source coding agent that lives in your terminal — built on LangChain Deep Agents. Ask it to fix a test, refactor a module, or carry a goal across turns until it is done.
English · 简体中文 · Docs · Changelog
Synapse is a terminal-first coding-agent runtime. Unlike a one-shot "run this and reply" prompt, it is designed for sessions that last more than a few turns: a responsive TUI that keeps the timeline, tool calls, and context usage visible; token-aware handling of tool output; and persistent goals that keep the agent working until the work is actually done.
Install
One command, no clone required:
uv tool install synapse-cli-agent
Windows users can also install via Scoop (single-file executable, auto-updated
through checkver): add the bucket, then install:
scoop bucket add synapse https://github.com/alex8224/synapse-agent
scoop install synapse
Or install directly from the manifest URL:
scoop install https://raw.githubusercontent.com/alex8224/synapse-agent/main/bucket/synapse.json
Then start the TUI from anywhere:
synapse tui -w .
To run the standalone foreground runtime daemon (S8), use the installed
synapse-runtime command or the module form. It prints one ready JSON metadata
line to stdout after binding; the token is generated in the private state
directory and is never printed.
synapse-runtime --state-dir ~/.synapse/runtime --host 127.0.0.1 --port 0
# or: python -m synapse.runtime.daemon --state-dir ~/.synapse/runtime --port 0
Use --token-file PATH to select a deployment-specific token file. The daemon
uses a held lock in the state directory, runs in the foreground, and stops on
SIGINT/SIGTERM. It does not provide install/start/stop/status controls and does
not migrate the CLI, TUI, or ACP consumers.
Web 控制台(React,正式宿主)
web/ 是 React + TypeScript 单页控制台。生产路径使用 synapse web-console
子命令(薄 aiohttp 进程:静态产物 + 配对 API + WebSocket 中继),
不依赖 Vite 开发中间件承担业务;Vite 只在开发时做静态热更新与受控代理。旧的
synapse-web(textual-serve 把 TUI 放进浏览器)入口保持不变、仍可用,与正式宿主
语义不同。
先同步依赖与前端产物再启动。一条命令即可:宿主默认会在该 --state-dir 下没有
运行中 daemon 时自己拉起一个(--port 0 由内核分配),并在自己退出时把它停掉;
已经有一个在跑则直接复用、绝不动它。想要「daemon 必须由我亲自启动」就加
--no-start-runtime;给了 --runtime-port 也视为「用那个 daemon」,不会另起。
uv sync # 源码检出必做:console script 由安装步骤生成
cd web && npm ci && npm run build && cd ..
# 一条命令:宿主按需拉起/复用 daemon(wheel 已内置前端产物)
synapse web-console --workspace . \
--state-dir ~/.synapse/runtime --port 8080
仍然可以把 daemon 当作独立服务来跑(服务化部署、或想让它跨控制台存活):先在一个 终端启动它,再启动宿主——宿主会发现它并复用。
synapse-runtime --state-dir ~/.synapse/runtime --host 127.0.0.1 --port 0
synapse web-console 与 synapse-runtime 都由安装步骤(uv sync)生成。未同步的源码检出里,
可以使用等价的模块形式:
uv run --no-sync python -m synapse.web_console.entry --workspace . \
--state-dir ~/.synapse/runtime --port 8080
宿主的自动拉起只覆盖它自己启动的 daemon:正常退出(Ctrl+C / SIGTERM)会连带停掉;
被强杀(任务管理器、taskkill /F)则 daemon 会留下来继续跑(它是服务,下次启动会被
复用而不是重复拉起)。两条命令行形式都是前台常驻进程,前台启动会占住调用它的终端:
交互使用各开一个终端;脚本/自动化
(含 agent)必须非阻塞启动——让子进程脱离调用方进程树,并把 stdout/stderr 重定向到
文件。PowerShell 下从 shell 直接 Start-Process -RedirectStandardOutput ... 会因
子进程继承 stdout 管道而卡住调用方(直到进程退出才返回),可行做法是 cmd.exe /c
包一层带重定向的脚本,再经 Win32_Process.Create 脱离启动:
$log = "$env:TEMP\synapse-webui"; New-Item -ItemType Directory -Force $log | Out-Null
# $log\start-console.cmd 内一行(@echo off 之后):
# "<repo>\.venv\Scripts\python.exe" -m synapse.web_console.entry --workspace "<repo>" ^
# --static-dir "<repo>\web\dist" --state-dir "%USERPROFILE%\.synapse\runtime" --port 8080 ^
# > "%TEMP%\synapse-webui\console.out" 2> "%TEMP%\synapse-webui\console.err"
Invoke-CimMethod -ClassName Win32_Process -MethodName Create -Arguments @{
CommandLine = "cmd.exe /c `"$log\start-console.cmd`""; CurrentDirectory = "$PWD" }
配对码随后从 $log\console.err 读取(前台运行时直接出现在宿主 stderr)。
Windows 上不必每次手敲命令:仓库自带 scripts/web-console-desktop.ps1,双击桌面图标
即可「一步到位」——脚本按需在后台拉起宿主(已在运行就只复用,不会重复启动),等端口
应答后打开控制台窗口。开窗优先级:已安装的 PWA(Start Menu 里带 --app-id 的
快捷方式,即真正的应用窗口——独立任务栏身份、窗口控件覆盖标题栏、任务栏跳转列表都在),
没有装 PWA 时用 Edge/Chrome 的 --app= 独立窗口,再退回默认浏览器。宿主与启动器日志、
宿主 PID 都在
%LOCALAPPDATA%\Synapse\web-console(console.out / console.err / launcher.log /
console.pid):
pwsh -NoProfile -File scripts\web-console-desktop.ps1 -InstallShortcut # 创建桌面快捷方式
pwsh -NoProfile -File scripts\web-console-desktop.ps1 -Restart # 停止并重启宿主,再开窗
pwsh -NoProfile -File scripts\web-console-desktop.ps1 -Stop # 停宿主(-WithDaemon 连 daemon 一起停)
# 同一脚本的另外两个桌面入口(-ShortcutAction start|restart|stop)
pwsh -NoProfile -File scripts\web-console-desktop.ps1 -InstallShortcut -ShortcutName "Synapse 控制台(重启)" -ShortcutAction restart
pwsh -NoProfile -File scripts\web-console-desktop.ps1 -InstallShortcut -ShortcutName "Synapse 控制台(停止)" -ShortcutAction stop
桌面图标默认带 --no-pairing:本机 loopback 单用户场景省掉每次输码;要保留配对码校验
就自己加 -Pairing,脚本会把配对码复制到剪贴板并弹窗提示。其它参数:-Port(默认
8080)、-Workspace(默认仓库根)、-DefaultBrowser(不开独立窗口)、-NoOpen、
-ShortcutAction。
-Restart 是「停宿主 → 等端口真正释放 → 重新拉起 → 开窗」,-Stop 只停宿主;两者都
不动 daemon——宿主自启的 daemon 作为服务继续运行、下次复用,正在跑的 agent turn
不会被重启打断;要连运行时一起重启就加 -WithDaemon。宿主换进程后已经开着的窗口需要
刷新(Ctrl+R)才会接上新宿主。
PWA 的 origin 在安装时就固定了(http://127.0.0.1:8080/ 或 http://localhost:8080/),
所以换 -Port 后 PWA 窗口可能连不上:这时加 -NoPwa 走 --app=,或用
-PwaShortcut "<Start Menu 里的 .lnk 路径>" 指定具体那个应用(同名装了多个时脚本会挑
不带 (1) 后缀的那个,并在 launcher.log 里记下实际用的是哪一个)。开窗前会校验应用
确实还装着——快捷方式目标 exe 存在,且该 --app-id 仍在对应浏览器 profile 的
Web Applications\Manifest Resources\<app-id> 下;残留的旧快捷方式会被跳过并记进
launcher.log,退回 --app= 或默认浏览器。
宿主启动时 stdout 恰好一行 JSON 元数据,配对码只出现在 stderr:
synapse web-console: pairing code XXXXXXXX (expires in 300s; open http://127.0.0.1:8080/ and enter it)
打开 http://127.0.0.1:8080/ 输入该 8 位配对码完成配对(单次使用,默认 300s
过期)。会话 cookie 默认只在 POST /api/pair 成功时签发;GET /api/bootstrap 已删除,
恒返回 405 且不签发 cookie。
本地调试(例如用 CDP/devtools 驱动控制台、不想每次重启宿主都重新输码)可以显式加
--no-pairing(默认关闭):宿主不创建、也不打印配对码,GET /api/session 会为任何同源
loopback 浏览器直接签发会话(200 + Set-Cookie,属性与配对签发一致),stderr 改打一条
WARNING pairing is disabled (--no-pairing) …,stdout 元数据行的 pairing_required 变为
false;POST /api/pair 在该模式下返回 400,跨站 POST 仍被 403 拒绝。它只保护到「同源
loopback 浏览器」这一层(Host/Origin/Sec-Fetch-Site 都是客户端可伪造的头,只算纵深
防御),没有任何持有证明,因此只用于本机调试,不要用在共享或对外暴露的控制台上。其余
边界不变:会话只在宿主进程内存、上限 8、默认 12h TTL,宿主重启即失效(该模式下由下一次
会话探测自动重新签发),/runtime-ws 中继仍要求有效会话 cookie,daemon bearer 依旧不出现
在任何 HTTP 响应、stdout 或中继帧里。
宿主默认且仅支持 loopback(127.0.0.1/localhost/::1,配置层强制),是 loopback
单用户工具,不是公网多租户安全产品。已建立的边界:浏览器始终拿不到 daemon
bearer token/env/model secret——宿主服务端持有 token 并以 Authorization: Bearer
连接 daemon,再中继 JSON-RPC 帧(唯一例外:指向非宿主自身项目的请求由宿主以 typed
not_found 拒绝、不进入 daemon,边界与残余风险见
docs/web-console/formal-host.md §4.1);URL 不再携带 token,前端也不再硬编码
token/project/workspace。配对码只存在于宿主进程内存与 stderr,不落盘、不进 stdout
JSON。/runtime-ws 中继在升级前要求 Host 与 Origin 精确匹配(含端口)+ 有效会话
cookie;浏览器断线不会取消 daemon 中仍在运行的 turn。静态服务带路径遍历/符号
链接逃逸防护、帧大小上界、SPA 回退与 no-store/asset immutable 缓存规则;中继出站
缓冲另有显式上界(64 帧 / 8 MiB,取先到者),超限即断开并计数。已配对浏览器可只读
GET /api/runtime-status 获取 daemon 端点、state dir 与 start synapse-runtime 提示
(响应不含 token;POST → 405)。
明确不承诺(详见 docs/web-console/formal-host.md §4、§9):Host/Origin/
Sec-Fetch-Site 都是客户端可伪造的头,只算纵深防御,默认模式下真正的门是配对码
(显式 --no-pairing 的豁免见上,它没有任何持有证明);本切片不支持
TLS/反向代理(外部端口 ≠ 绑定端口);会话不持久化,宿主重启后需重新配对
(--no-pairing 下由下一次会话探测自动重新签发,仍非持久化);
会话 TTL 到期不主动断开已建立的中继;中继缓冲上界是进程内上界(不含
aiohttp/内核 socket 缓冲,不等于进程 RSS 上界);跨项目拒绝对不可读/非文本帧
fail-open,且白名单与 protocol.decode_params 无源码级绑定,安全性依赖 daemon 的
严格校验;历史投影与实时 broker 之间不承诺跨存储原子一致(见 docs/sessions.md)。
开发模式(仅热更新与受控代理,不经中间件读 token):
synapse web-console --workspace . --static-dir web/dist --port 8080 &
cd web && npm run dev # 打开 http://127.0.0.1:5173
Vite 只把 /api 与 /runtime-ws 代理到宿主,并把转发请求的 Origin 重写为宿主
origin;它不直连 daemon、不注入任何凭据、不读 token 文件,server.host 保持
loopback 默认值,因此不构成认证旁路,宿主也不因 dev 放宽任何校验。dev 下同样需要
配对码(从宿主 stderr 读取;是否豁免完全由宿主的 --pairing/--no-pairing 决定)。
完整命令与参数见 docs/web-console/formal-host.md。
控制台里的每个弹框/浮层都能只用键盘操作(F2 模型选择、F5 MCP 面板、F6 目标、
输入框左侧 + 的添加项目、顶栏分支 chip 的 Git Explorer、设置对话框):打开后焦点
立刻进入框内(模型选择器落在过滤输入框、推理等级落在当前等级,目录 / Git 变更 /
MCP 服务器列表落在第一行),↑/↓ 按阅读顺序在框内控件之间移动并首尾相接,
Enter/Space 触发聚焦项,Esc 关闭,关闭后焦点回到触发它的那个控件。
回归验收:cd web && node tests/dialogKeyboardNav.verify.ts。
控制台也能作为应用安装(PWA,web/public/manifest.webmanifest):Chrome 用地址栏
右侧的安装图标,Edge 用「应用 → 安装此站点为应用」,即可得到独立窗口的 Synapse;
任务栏右键另有「新建会话」快捷方式。安装窗口的标题栏由控制台自己的顶栏接管
(display_override: window-controls-overlay,浏览器只保留窗口按钮;不支持该模式的浏览器
退回普通应用窗口)。安装只是外壳——宿主仍然要在跑
(synapse web-console 及其 daemon),页面关掉后不会有后台推送,宿主重启后也要重新配对。
宿主若显式用 --no-pairing,则重启后由下一次会话探测自动重新签发会话(仍非持久化)。
细节与边界见 web/README.md「可安装应用(PWA)」。
Open a session in any registered project from the global catalog:
synapse --session <project_id>:<thread_id> # global session reference
synapse --project <ref> # project by id prefix, name, or path
Inside the TUI, click the topbar ≡ (or the workspace label) to open the
floating project drawer: it lists every registered project, groups their
sessions, marks live runtime status, and lets you switch sessions in place or
jump to another project (the TUI restarts into that project).
Press Ctrl+Tab for a global recent-sessions switcher: it lists the most
recently changed sessions across every registered project (from the user-layer
catalog), marks each one's live runtime status — running / queued /
idle / cold, etc. — lets you Tab / Shift+Tab through them and Enter
to switch. The old session keeps running in the background and is never
cancelled. Some terminals never forward Ctrl+Tab to the app; Ctrl+O is a
drop-in alternative.
/new and switches to cold sessions prepare an independent agent in a background
worker, leaving the current session visible until preparation succeeds. While
preparing, submitted prompt text stays in the input instead of being sent to the
old session; submit it again after the switch. A failed preparation leaves the
current session unchanged. Mid-turn steer notifications do not wait for UI repaint,
and streaming updates are coalesced into bounded UI batches.
Or run from source
git clone https://github.com/alex8224/synapse-agent.git
cd synapse-agent
uv sync
# Windows venv entry
.\.venv\Scripts\synapse.exe tui -w .
# Or module entry
uv run python -m synapse tui -w .
语音输入
控制台输入区控制行有一个麦克风按钮,把说的话变成输入框里的文字。识别引擎由设置决定:
| 引擎 | 启用方式 | 特点 |
|---|---|---|
browser(默认) |
无需配置 | 浏览器自带的 Web Speech API(Chrome / Edge)。免费、零安装,但音频要经浏览器厂商服务器识别,中文一般、英文术语容易丢 |
local |
STT_ENGINE=local + uv sync --extra stt-local + 模型文件 |
本地离线 ONNX 引擎(synapse.stt):免费、完全离线,中文明显更好(实测把 runtime 的 attach 接口 完整认对)。代价是首次听写要等模型加载(这台机器上约 1 分钟,之后常驻) |
doubao |
桌面端 + 在「设置 → 语音输入」里填豆包 API Key | 火山引擎豆包大模型流式识别:中文最强、支持热词与上下文;按小时计费,音频会上传到服务商。只在桌面端可用——它靠自定义握手请求头鉴权,浏览器无法设置 |
local 引擎是两遍解码:流式模型边说边出字,显示在输入框下方的实时字幕里;一句话说完后,
离线模型重新解码该句并纠正,纠正后的文字才落进输入框。所以字幕里的字是临时的,输入框里的
是定稿。
doubao 引擎的连接由桌面端自己建立,不经 daemon 中转:API Key 存在
~/.synapse/stt.json(与 settings.json 分开,因为后者被文档定义为非密文件),桌面端直接读它、
只在桌面进程里使用,从不回传给控制台页面。在纯浏览器里选择云端引擎会回退到浏览器内置识别,
并在输入卡片里说明原因。
延迟上做了两处实测调优(同一段 10.9 s 真实语音、同一服务、5 轮以上交错试验):分片从 200 ms 降到 厂商建议下限 100 ms,首字提前约 150 ms、句中词提前约 175 ms;收句静音窗口从厂商默认的 800 ms 降到 400 ms,句子落进输入框提前 400 ms,且句子边界不变。
中转本身不是延迟来源(回环一跳亚毫秒)。"停止说话 → 文字进输入框"现在的构成是:收句窗口
400 ms + 厂商流式收句约 350 ms + 第二遍解码约 270 ms。第二遍可以一行关掉
(CloudTuning::nonstream)再省 270 ms,但它是厂商的准确率特性,且实测在这段音频上反而把
runtime/attach 改成大写(流式那遍是对的),一段样本不足以推翻它,所以默认保留。剩下的大头是
厂商攒够约 700 ms 音频才吐第一个字——文档里那个"首字加速"旋钮我试了,见下。
文档里还有两档我们没有默认启用(CloudTuning 里都是一行):end_window_size 官方最小值
200 ms(实测再省 267 ms),和关掉第二遍(实测再省 200–300 ms,且这段音频上文本反而更对)。
两者都是准确率权衡:这段样本句内最长静音只有 150 ms,结构上测不出 200 ms 会不会在从句停顿处
断句,所以默认停在 400。厂商的"首字加速"(enable_accelerate_text / accelerate_score)我也实测了,
在这段音频上没有任何效果——但厂商对不支持的参数会静默忽略,所以那可能是"没生效"而非"没用"。
模型只在第一次用到时构建(这台机器上约 1 分钟),而构建由真正需要它的那一次操作触发:点麦克风, 或在「设置 → 语音输入」里按「加载模型」。构建期间麦克风旁显示「正在加载语音模型(首次约 1 分钟)」、 按钮不可点,避免把这一分钟藏进一次听写里。打开控制台不会自动构建——否则在你还没决定用哪个引擎 之前就先跑起一分钟的本地模型,想改成别的引擎反而被拖住。构建完成后一直常驻,后续听写是秒级。
切换引擎不用改文件:控制台的「设置 → 语音输入」里可以直接选浏览器内置、本地引擎或云端引擎
(以及模型目录和 API Key),保存会写入 ~/.synapse/settings.json 并立即生效——不需要重启,也不
需要刷新页面。上面那两个环境变量是给脚本化/无人值守场景用的,效果相同。
模型放在 ~/.synapse/stt/models(STT_MODEL_DIR 可改),缺失时控制台会说明缺哪些文件并
回退到浏览器引擎。变量说明见 docs/config.md。
Use it
Type it like you would ask a colleague:
- "Fix the failing test in
tests/test_backends.py" - "Refactor this module and add type annotations"
- "Review the latest commit and suggest improvements"
- "Keep working on this goal until it is done" (via
/goal)
Three ways to talk to it:
| Mode | What it does |
|---|---|
| TUI | Full terminal UI: timeline, turn rail, tool groups, context usage, themes |
| Chat | Plain interactive REPL — synapse chat -w . |
| Run | One-shot task that prints the answer — synapse run "summarize this repo" -w . |
Why it is built for long sessions
- Long-running goals —
/goal <objective>survives turn boundaries, tracks tokens and elapsed time, and steers the next turn automatically until the goal is completed, paused, blocked, or budget-limited. - Token-aware tool output — search results, logs, diffs, JSON, and code are classified and compressed before they re-enter the model context; large originals stay recoverable through references.
- Managed long context — automatic summarization and
/compactkeep sessions inside the model window, with occupancy and savings visible in the TUI. - Structured system prompt — the prompt is assembled from named, cache-hinted sections whose stable instructions stay byte-identical.
AGENT_ENABLE_PROMPT_CACHE_BOUNDARY(off by default) additionally appends the per-build environment, git snapshot, and current date as dynamic context and marks the stable/dynamic seam with a prompt-cache breakpoint, so the volatile tail never invalidates the cached prefix. - Direct Codex OAuth — sign in with the Codex-compatible browser flow or import an existing Codex grant. No API key required; tokens refresh automatically.
- Your model, your choice — OpenAI-compatible providers via
models.jsonprofiles (OpenAI, DeepSeek, local gateways, and more), including a persistent WebSocket mode. - MCP built in — attach MCP servers and their tools appear in the agent automatically.
- Programmatic tool calling (PTC) — in
both/codemode the model can orchestrate many tool calls from onerun_code(code, intent)block inside a one-shot local subprocess (ideal for batching MCP calls). Off by default (native); not a security sandbox. - Sessions that resume — SQLite checkpoints, a global project catalog across all your projects, and a lightweight paged transcript so even huge sessions reopen fast.
- Memory and skills —
AGENTS.mdmemory plus Agent Skills (skills/**) that load only when relevant. - Sub-agents — built-in
researcher,tester, andreviewerroles for parallel delegation, plus user-defined subagents from.synapse/agents/*.md. Each runs in its own context against the shared workspace, and the global tool policy (minimal filesystem / read-only) is always enforced. - Approvals when you want them — optional human-in-the-loop (
--require-approval) and safety profiles.
Slash commands
Type /help in the TUI for the full reference. The essentials:
| Command | What it does |
|---|---|
/goal <objective> |
Set a long-running goal that auto-continues across turns |
/goal pause · /goal resume · /goal clear |
Manage the active goal |
/new · /switch <id> · /sessions |
Create, switch, and list sessions |
/export [md|json] |
Export the transcript to a file |
/model <provider:model> |
Switch models at runtime |
/model manage · /model import-codex · /model providers |
Manage profiles, import Codex config, list providers |
/fast [on|off|status] |
Toggle the Codex Fast tier (OAuth profiles) |
/mcp list · /mcp reload |
Manage MCP servers |
/theme <name> |
Switch UI themes |
/compact |
Force context compaction |
/context |
Show context usage stats |
/safety <profile> |
Switch safety profiles |
/approve · /reject |
Human-in-the-loop decisions |
TUI error diagnostics
Turn and approval-resume errors show a nonempty summary (including the exception
type when its message is empty) and an Error log: path. Failure diagnostics are
enabled by default and written lazily to the turn's own workspace:
.synapse/logs/errors-<pid>.log. Each process rotates its file at 1 MiB with three
backups. If the file cannot be written, the TUI explicitly reports that instead
of hiding the original error.
The JSON-lines log keeps timestamps, operation/session/turn identifiers,
exception types, numeric error codes and stack locations. It deliberately omits
raw exception messages, source lines, locals, prompts, tool output and settings;
the on-screen summary redacts common credential formats. Runtime failures are
recorded before conversion to result data, when the traceback is still available.
See TUI error details and logs for the PowerShell
tail command and retention details. Restart the TUI after updating; earlier
unlogged failures cannot be recovered from an ERROR: screenshot.
ACP v1 adapter
The package also installs the standalone synapse-acp stdio entry point. It
uses the locked agent-client-protocol==0.12.0 dependency and keeps stdout
reserved for ACP JSON-RPC; diagnostics go to stderr.
# Run from an installed package
synapse-acp
# Run from a source checkout
uv run synapse-acp
For a client such as Zed, configure the ACP agent as a subprocess whose command
is synapse-acp (or the absolute path to that executable). The adapter accepts
absolute cwd values and session-scoped MCP servers. MCP credentials are used
only for the live session and are not written to the ACP catalog or transcript.
The published capability set currently covers text prompts, image prompts,
permission/HITL, session load/list/close/delete/resume, session-local
mode/config, HTTP/SSE MCP configuration, model selection via
providers/list/providers/set, and capability-gated Client filesystem/
terminal tools. Thinking level maps directly to Synapse levels (off,
minimal, low, medium, high, max). Audio, embedded-context,
authentication/logout, elicitation, NES, and document-sync capabilities are
not advertised until their runtime semantics and interoperability tests exist.
Troubleshooting:
- stdout is reserved for ACP JSON-RPC; any log line on stdout breaks the protocol. Logs and startup traces go to stderr.
cwdandadditionalDirectoriesmust be absolute paths; relative paths are rejected with an ACP error.- Session metadata lives in
~/.synapse/acp-sessions.sqlite; MCP credentials and headers are never persisted there. - If a client reports "Method not found", verify it negotiates the same ACP
protocol version and does not rely on
not_targetmethods (auth, providers, elicitation, NES, document sync).
Pick a model
Synapse works with any OpenAI-compatible endpoint. Configure profiles in ~/.synapse/models.json (or <workspace>/.synapse/models.json):
{
"default": "deepseek",
"models": {
"deepseek": {
"model": "openai:deepseek-v4-pro",
"api_key": "sk-...",
"base_url": "http://127.0.0.1:3000/v1",
"thinking": "high"
}
}
}
For a zero-config Codex experience, use the OAuth profile — see Models.
Documentation
| Quickstart | First run, CLI reference, common workflows |
| Configuration | Layered settings, environment variables, paths |
| Models | Provider profiles, OAuth, Fast tier, WebSocket mode |
| MCP | Attaching and managing MCP servers |
| PTC | Programmatic tool calling: run_code, modes, limits, safety |
| Sessions | Checkpoints, resume, transcript paging |
| Skills | Bundled skills and the Agent Skills format |
| Permissions | Read-only mode and approval flows |
| Install | All installation methods |
| ACP adapter | ACP v1 setup, capability matrix, limitations, and verification status |
| Runtime contract | v1 wire/DTO contract: generated manifest and TS types, compatibility rules, gates |
The runtime contract (wire method table, event kinds, DTO schemas) is generated
from src/synapse/runtime/service/contract_registry.py into
src/synapse/runtime/service/contract_manifest.json and
web/src/runtime-client/contract.generated.ts. Do not hand-edit either artifact:
uv run --no-sync python scripts/export_contract_manifest.py # regenerate
uv run --no-sync python scripts/export_contract_manifest.py --check # verify committed output
Repository layout
src/synapse/ Agent assembly, commands, runtime, sessions, TUI, integrations
rust/ Optional native compression cores (PyO3)
docs/ User documentation
tests/ Python test suite
scripts/ Install / release helpers
License
Apache License 2.0 — see LICENSE. Third-party dependencies and Rust subcomponents may carry their own licenses; retain the corresponding license and NOTICE files when distributing.
Metadata
Release files for synapse-cli-agent 0.1.58
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| synapse_cli_agent-0.1.58.tar.gz | 11.7 MB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| synapse_cli_agent-0.1.58-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 16.0 MB
Release files / synapse_cli_agent-0.1.58.tar.gz
| Download URL | synapse_cli_agent-0.1.58.tar.gz |
|---|---|
| Size | 11.7 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
c8cec87acd1e1c3e5500a4a1b21eb3a241be872c04490922e531e68d1c4389a7
|
|
BLAKE2b-256 checksum How to use checksums |
5ab42a22017af081d6e7d82d7f1c0f383ffc298a9bf706638a3813a8d9a93aa7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 12, 2026.
Transparency logRelease files / synapse_cli_agent-0.1.58-py3-none-any.whl
| Download URL | synapse_cli_agent-0.1.58-py3-none-any.whl |
|---|---|
| Size | 4.3 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
8ad22f7ab6500390c15e374113400cb5f74f6125e3afacbfd2746b76124ed4aa
|
|
BLAKE2b-256 checksum How to use checksums |
a94365bcb74967b9e777447f03a3abe142761219609f377bf1d8bf709cd48396
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 12, 2026.
Transparency log