Knowledge 功能需额外执行 synapse upgrade --with knowledge 安装 Docling 和 CPU 版 PyTorch；Docker 部署时后端端口 8765 只应绑定本机，切勿发布到 0.0.0.0；从 1.x 升级前不要先运行 2.0 的 synapse uninstall，否则会连同备份一起删除 ~/.synapse。
synapse-ai
synapseorch-aiAGPL-3.0活跃维护开源 AI Agent 编排平台，用 DAG 把多个智能体、MCP 工具和任意 LLM（本地、云端或 CLI）串起来，让 Agent 真正能干活。
Build AI agents that actually do things. Synapse is an open-source platform for creating, connecting, and orchestrating AI agents powered by any LLM — local, cloud or CLIs.
v2.0.1发布于 2026-11-09 14:02 UTC338star63fork同步于 2026-11-10 18:02 UTC
curl -LsSf https://synapseorch.com/install.sh | bash
npm install -g synapse-orch-ai && synapse 硬件要求按 README 说明
README 未说明显存或 GPU 要求；安装器自带 Python，必要时自动安装 Node.js，可在笔记本或服务器上运行，也支持 Docker 部署。
有 Docker 镜像
开源 AI 智能体编排平台，用一句话描述任务，生成可审阅的工作流并按计划自动执行
开源 AI Agent 编排平台，用 DAG 把多个智能体、MCP 工具和任意 LLM（本地、云端或 CLI）串起来，让 Agent 真正能干活。
核心能力
- 用自然语言描述任务，自动生成带步骤和成本的计划
- 在工作流中插入人工签核步骤，支持 Slack、Telegram 等渠道审批
- 每个步骤可独立选择模型，支持 23+ 提供商和本地 Ollama
- 提供 CLI、REST API、MCP 服务和 OpenAI 兼容端点
- 按运行设置硬性花费上限，并展示每步成本
容易踩的坑
内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README
- v2.0.1最新正式版2026-11-09 · 2026-11-09 14:02 UTC
Synapse 2.0.1
A small release on top of 2.0.0: clearer answers when a build stops, models that are easier to find, and Synapse AI reaching for the apps you can already connect.
Fixed
- A build that stops says why. When Synapse AI stops partway through building something, the chat now gives the reason the step recorded, for example that the model ran out of credit, instead of "it stopped before anything was saved".
- Models are easier to find. Every model picker shows each model's provider icon, and typing a provider's name ("anthropic", "openai") finds its models. The built-in list of Claude models is current: Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1, Opus 4.8, Sonnet 4.6 and Haiku 4.5.
- Apps come first. When a tool you ask for belongs to a service in the Apps catalogue (GitHub, for example), Synapse AI plans it on that app and connects it on its card after you say yes, rather than building its own API tool for it.
- An API address with a fixed query works. An endpoint written with its own query, such as GitHub's
/pulls?state=open, keeps its API, and the query stays on the tool's address. An API that wants its key written into the address gets the plain answer that Synapse AI can't keep a key there safely. - The PyPI and npm pages link to synapseorch.com, with the docs beside it.
Install
macOS and Linux
curl -LsSf https://synapseorch.com/install.sh | bashWindows (PowerShell)
irm https://synapseorch.com/install.ps1 | iexnpm install -g synapse-orch-ai && synapse pip install synapse-orch-ai && synapse # Python 3.11–3.14 and Node.js 20.9+For Docker, see Docker: the image is
synapseorchai/synapse-ai:2.0.1, also:latest.Upgrading
- From 2.0.0: run
synapse upgrade. Your data stays where it is, and the database is backed up first. - From 1.x: run
synapse upgradeon your 1.x install. It moves you to 2.0 the same way 2.0.0 did; read the 2.0.0 notes for what changes.
Full changelog: v2.0.0...v2.0.1 · Documentation
- v2.0.02026-11-08 · 2026-11-08 19:32 UTC
查看更新说明
SynapseOrch 2.0
Describe a job in plain words. Synapse AI shows you the plan, builds the agents and the workflow once you approve, and runs it exactly as approved.
https://github.com/user-attachments/assets/26147f4f-804e-4086-8fa2-658af526783f
2.0 is the biggest release of SynapseOrch so far, and the first one we'd call complete. It adds:
- a builder you talk to;
- a run page that tells you what happened;
- a browser your agents can use, with you able to take over;
- answers from your own files;
- agents on your website and in your channels;
- a new installer.
It's open source (AGPL-3.0). It runs on your own machine or server, and works with the models you already pay for. That includes the Claude Code and Codex CLIs, so there are no extra API keys.
Run it yourself:
curl -LsSf https://synapseorch.com/install.sh | bashOr try SynapseOrch Cloud with nothing to install. Already on 1.x? Run
synapse upgrade(see Upgrading from 1.x).What's new
Synapse AI builds it with you
Tell Synapse AI the job, for example "every morning, pull yesterday's orders from our shop API and post a summary to Slack".
- It researches. It reads the sites and API docs you name.
- It shows the plan in plain words.
- It builds the agents, tools and workflow, only after you approve.
An API it finds is tried before it's kept. One that needs no key is made without asking you for one, and if an API needs you to sign in, Synapse AI says so. Later, ask it what ran, what failed and why, and what it cost.
https://github.com/user-attachments/assets/131484b5-0f54-4936-b06e-3a5475aa3de3
Chat opens on Synapse AI on every kind of install: the installer, npm, pip and Docker.
Every change is shown before it happens
Ask for a change and you get a review card: each change beside what was there, and code as a diff. A fix to a Python tool is tested with no internet access before you see it.
https://github.com/user-attachments/assets/4ce3ca5d-c81a-4c1b-976a-c3ac860d0931
Paste a
curl, an OpenAPI spec or a docs link, and Synapse AI makes the HTTP tool. Any key goes to a Connect card, never into the tool.https://github.com/user-attachments/assets/1c0115ce-70c6-4ba7-944a-c49cbb38760f
Synapse AI is a workflow you can open
Synapse AI isn't a black box. It's an ordinary workflow with ordinary agents, built from the same steps you use. See how I'm built opens it on a canvas: where it starts, where it stops to ask you, and which agent does what. Duplicate it to make your own.
https://github.com/user-attachments/assets/35aa5fac-5530-4bd5-ad8c-a2c9a6b8cecc
A run page that shows what happened
Every run has its own page. The result is at the top, followed by:
- the path the run took;
- what each step produced;
- how long each step took and what it cost.
A failed step's card says what went wrong, in plain words. Runs of a workflow you later deleted stay in Runs.
https://github.com/user-attachments/assets/7c05c394-2cf9-4a13-a03c-5d0ac09aff39
Sign-offs where you want them
A step can stop and wait for a person: an approval, or a short form such as an amount or a reason. Answer it in the app or from a chat channel, and the run carries on with your answer. Risky tool calls ask first.
https://github.com/user-attachments/assets/19505fbb-8909-4ced-9114-c872085c6201
A browser for your agents, with you at the wheel
Browser agents work real sites in SynapseOrch's own browser. You watch every action in a live pane, and you can take over at any point and hand back.
Sign in to a site once and your agents start signed in. Sessions are kept; passwords never are. Record a task once, and a Browser step repeats it with no model calls.
https://github.com/user-attachments/assets/e7b436bf-3ad9-4057-8e09-7f3ac3e09118
- v2.0.0-rc3pre-release2026-11-07 · 2026-11-07 22:23 UTC
查看更新说明
Synapse 2.0, release candidate 3
Describe a job in plain words. Synapse AI builds the agents and the workflow, you approve every step, and it runs exactly as approved.
This is the third release candidate of Synapse 2.0, a new major version: a new builder, a new run page, a built-in browser, a new installer and one store for everything. It's open source (AGPL-3.0), runs on your own machine or server, and works with the models you already pay for.
A release candidate is for trying before 2.0.0. It's what we intend to ship, and we'd like to hear what breaks: open an issue. It isn't the latest release. The usual install line, npm's default, Docker's
:latestand 1.x'ssynapse upgradeall stay on 1.9.3 until 2.0.0 is out. To try it, use the lines under Install the release candidate.Fixed since rc2
- Turn a schedule on or pause it from its row. Schedules has an On / Paused switch in each row, and Turn on / Pause in the row's menu, where before you opened the schedule, ticked Enabled and saved. Pausing is one click; turning on first shows the message each run starts with and when it first runs.
- Synapse AI turns on a schedule it made when you say yes. A schedule it creates still starts paused. It now shows you what each run will do and when the first one is, and your yes turns it on, where before it sent you to the Schedules screen. A schedule turned on after its first run's time has passed works out its next run from now instead of keeping the missed one.
- Over MCP,
turn_on_scheduleandpause_schedulereplaceset_schedule_enabled.turn_on_scheduleis marked so your MCP client asks you before it runs;pause_scheduleisn't. - A knowledge base is listed as soon as you make it. It could take several seconds to appear, under "No knowledge bases yet", while the knowledge service woke up.
Fixed since rc1
- Synapse AI keeps building when a model wraps its plan. Some models (seen with Claude Haiku 4.5) sent the plan in the shape Synapse AI uses for drafts, and the build stopped before anything was saved. That plan is now read as the plan.
- A pip install keeps its access link and sign-in. The first
synapse startafterpip installcouldn't save them, sosynapse access-linkprinted a link that didn't open Synapse, and the next restart signed everyone out. - A failed install says how to try the same release again. Its "Try again" line pointed at the latest release, which is 1.9.3 for a release candidate and has no setup script there.
synapse upgradefinds a release candidate's installer. It asked GitHub forv2.0.0rc1when the release isv2.0.0-rc1.- The installer moves past a download server that doesn't answer. It gives each address 10 seconds, tries the next, and says when it's still trying, where before it could sit at "Getting the Synapse installer…" for minutes.
- Upgrading from 1.x counts only your own work. The upgrade question counted 1.x's built-in agents and its builder's workflow as yours ("6 agents" when you had made none), and now says "workflows", as 2.0 does.
On rc1 or rc2? Run rc3's install line below, or
npm install -g synapse-orch-ai@next, orpip install synapse-orch-ai==2.0.0rc3. Your data stays where it is. rc1's ownsynapse upgradecan't fetch a later release candidate (the fix above shipped in rc2), so on rc1 use one of these lines.What's new
Synapse AI builds it with you
Tell Synapse AI the job, say "every morning, pull yesterday's orders from our shop API and post a summary to Slack". It reads the sites and API docs you name, shows a plan in plain words, and builds the agents, tools and workflow only after you approve it. An API it finds is tried before it's kept, and one that needs no key is made without asking you for one. If an API needs you to sign in, Syna
- v2.0.0-rc2pre-release2026-11-07 · 2026-11-07 06:47 UTC
查看更新说明
Synapse 2.0, release candidate 2
Describe a job in plain words. Synapse AI builds the agents and the workflow, you approve every step, and it runs exactly as approved.
This is the second release candidate of Synapse 2.0, a new major version: a new builder, a new run page, a built-in browser, a new installer and one store for everything. It's open source (AGPL-3.0), runs on your own machine or server, and works with the models you already pay for.
A release candidate is for trying before 2.0.0. It's what we intend to ship, and we'd like to hear what breaks: open an issue. It isn't the latest release. The usual install line, npm's default, Docker's
:latestand 1.x'ssynapse upgradeall stay on 1.9.3 until 2.0.0 is out. To try it, use the lines under Install the release candidate.Fixed since rc1
- Synapse AI keeps building when a model wraps its plan. Some models (seen with Claude Haiku 4.5) sent the plan in the shape Synapse AI uses for drafts, and the build stopped before anything was saved. That plan is now read as the plan.
- A pip install keeps its access link and sign-in. The first
synapse startafterpip installcouldn't save them, sosynapse access-linkprinted a link that didn't open Synapse, and the next restart signed everyone out. - A failed install says how to try the same release again. Its "Try again" line pointed at the latest release, which is 1.9.3 for a release candidate and has no setup script there.
synapse upgradefinds a release candidate's installer. It asked GitHub forv2.0.0rc1when the release isv2.0.0-rc1.- The installer moves past a download server that doesn't answer. It gives each address 10 seconds, tries the next, and says when it's still trying, where before it could sit at "Getting the Synapse installer…" for minutes.
- Upgrading from 1.x counts only your own work. The upgrade question counted 1.x's built-in agents and its builder's workflow as yours ("6 agents" when you had made none), and now says "workflows", as 2.0 does.
On rc1? Run rc2's install line below, or
npm install -g synapse-orch-ai@next, orpip install synapse-orch-ai==2.0.0rc2. Your data stays where it is. rc1's ownsynapse upgradecan't fetch rc2 (the fix above ships in rc2), so use one of these lines.What's new
Synapse AI builds it with you
Tell Synapse AI the job, say "every morning, pull yesterday's orders from our shop API and post a summary to Slack". It reads the sites and API docs you name, shows a plan in plain words, and builds the agents, tools and workflow only after you approve it. An API it finds is tried before it's kept, and one that needs no key is made without asking you for one. If an API needs you to sign in, Synapse AI says so. Later, ask it what ran, what failed and why, and what it cost.
Chat opens on Synapse AI, on every kind of install: the installer, npm, pip and Docker.
Every change is shown before it happens
Ask for a change and you get a review card: each change beside what was there, code as a diff. Paste a
curl, an OpenAPI spec or a docs link, and it makes the HTTP tool, with any key moved to a Connect card, never into the tool. A fix to a Python tool is tested with no internet access before you see it.Synapse AI is a workflow you can open
Synapse AI isn't a black box: it's an ordinary workflow with ordinary agents, built from the same steps you use. See how I'm built opens it on a canvas, showing where it starts, where it stops to ask you and which agent does what. Duplicate it to make your own.
A run page that shows what happened
Every run has its own page: the path it took, what each step produced, how long it took and what it cos
- v2.0.0-rc1pre-release2026-11-07 · 2026-11-07 04:06 UTC
查看更新说明
Synapse 2.0, release candidate 1
Describe a job in plain words. Synapse AI builds the agents and the workflow, you approve every step, and it runs exactly as approved.
This is the first release candidate of Synapse 2.0, a new major version: a new builder, a new run page, a built-in browser, a new installer and one store for everything. It's open source (AGPL-3.0), runs on your own machine or server, and works with the models you already pay for.
A release candidate is for trying before 2.0.0. It's what we intend to ship, and we'd like to hear what breaks: open an issue. It isn't the latest release. The usual install line, npm's default, Docker's
:latestand 1.x'ssynapse upgradeall stay on 1.9.3 until 2.0.0 is out. To try it, use the lines under Install the release candidate.What's new
Synapse AI builds it with you
Tell Synapse AI the job, say "every morning, pull yesterday's orders from our shop API and post a summary to Slack". It reads the sites and API docs you name, shows a plan in plain words, and builds the agents, tools and workflow only after you approve it. An API it finds is tried before it's kept, and one that needs no key is made without asking you for one. If an API needs you to sign in, Synapse AI says so. Later, ask it what ran, what failed and why, and what it cost.
Chat opens on Synapse AI, on every kind of install: the installer, npm, pip and Docker.
Every change is shown before it happens
Ask for a change and you get a review card: each change beside what was there, code as a diff. Paste a
curl, an OpenAPI spec or a docs link, and it makes the HTTP tool, with any key moved to a Connect card, never into the tool. A fix to a Python tool is tested with no internet access before you see it.Synapse AI is a workflow you can open
Synapse AI isn't a black box: it's an ordinary workflow with ordinary agents, built from the same steps you use. See how I'm built opens it on a canvas, showing where it starts, where it stops to ask you and which agent does what. Duplicate it to make your own.
A run page that shows what happened
Every run has its own page: the path it took, what each step produced, how long it took and what it cost, with the result at the top. A failed step's card says what went wrong. Logs names each run by what it is, in plain words. Runs of a workflow you later deleted stay in Runs.
Sign-offs where you want them
A step can stop and wait for a person: an approval, or a short form such as an amount or a reason. You answer it in the app or from a chat channel, and the run carries on with your answer. A Yes / No question takes a plain yes or no, typed or replied. When Synapse AI builds a sign-off with an approved path and a declined one, each ends at its own End, so an approved run never shows the declined End. Risky tool calls ask first.
A built-in browser for your agents
Browser agents work real sites in Synapse's own browser, and you watch every action in a live pane. You can take over at any point and hand back. Your own browser, for signing in or recording a task, opens as a panel on the right that you can widen. Its ✕ asks before closing when there's something to lose, a sign-in you've started or a recording with steps; one you haven't touched just closes. In every pane, what the agent did folds to one line under the page and opens to the full list; a browser task you ask for brings back a pane you'd hidden. When you sign in to a site for your agents, the browser opens on that site. Sign-in sessions are kept for your agents; passwords never are. Record a task once and a Browser step repeats it with no model
用户点评
登录用户的真实使用记录，可按版本和硬件筛选共 0 条
还没有人点评。第一个说出真实体验的人最有价值。
Hand AI the work your team repeats. Approve what goes out.
Describe a job in plain words. SynapseOrch turns it into steps you can read, runs them exactly as you approved, every time, and asks before anything goes out. Any model, your own tools and documents, on your laptop or your servers. Free and open source under AGPL-3.0, with no account and no licence check.
Describe it. Approve it. Rely on it.
- Describe it. Tell Synapse AI the job. It writes the steps, picks a model for each and marks where you sign off, then shows you the plan before anything is built.
- Approve it. Read every step. When a run reaches yours, it waits for your yes — in the app, or in Slack, Telegram or another chat channel.
- Rely on it. It runs on schedule exactly as you approved, and shows what each step did and what it cost. A run never passes the spending limit you set.
What's inside
| Synapse AI | An assistant that builds and changes agents and workflows from a sentence, and answers questions about your workspace. |
| Workflows | Steps you can read on a canvas: agents, models, tools, code, branches, loops, parallel paths, sign-offs and other workflows. Runs follow the path you drew. |
| Agents and teams | Agents with their own model, tools, skills and memory. A team agent splits work among its members; any agent can start short-lived helpers. |
| Sign-offs | A step that waits for a person — answered in the app or in a chat channel. Approvals for risky tool calls, too. |
| Knowledge | Your files and web pages, indexed; answers cite the passage they used. (Optional part — see below.) |
| Boards | A shared to-do board for people and agents, with each card's history. (Off until you turn it on in Settings.) |
| Apps and tools | MCP servers (local or remote, OAuth included), REST and OpenAPI, Python functions, SQL databases, web fetch and scraping. |
| Channels | Your agents on Slack, Discord, Telegram, Microsoft Teams, WhatsApp, a website chat widget, or a phone number (Twilio). |
| Any model | Your own keys for 23+ providers, any OpenAI-compatible endpoint, local models through Ollama, or the Claude Code, Codex, Gemini and Copilot CLIs you already use. A different model at every step. |
| Costs you can read | An estimate before a run, what each step cost after it, and a hard limit per run. |
| Use it from anywhere | A command line, a versioned REST API, an MCP server for Claude, ChatGPT and Cursor, and an OpenAI-compatible endpoint where every agent answers as a model. |
| Scale mode | Redis queue, a worker fleet, Postgres checkpoints and S3 storage, for when one process is not enough. |
Install
macOS / Linux
curl -LsSf https://synapseorch.com/install.sh | bash
Windows (PowerShell)
irm https://synapseorch.com/install.ps1 | iex
Nothing needs to be installed first: the installer brings its own Python and, if yours is missing or too old,
Node.js — no sudo. It asks which AI to use (and whether to start on login) while it installs, and ends with the link.
Your data and settings live in ~/.synapse. Add --yes (curl … | bash -s -- --yes) to take every default.
The short addresses redirect to the scripts on the latest GitHub release: setup.sh
and setup.ps1.
npm install -g synapse-orch-ai && synapse # the first run uses the same installer
pip install synapse-orch-ai && synapse # Python 3.11–3.14 and Node.js 20.9+
docker run -d --name synapse -p 3000:3000 -p 127.0.0.1:8765:8765 \
-v synapse-data:/data \
-v /var/run/docker.sock:/var/run/docker.sock \
synapseorchai/synapse-ai:latest
Then open http://localhost:3000. The backend's port (8765) is published on this machine only: it serves an
internal API, so never publish it on 0.0.0.0.
Optional parts. Knowledge brings Docling and a CPU build of PyTorch, so it is added only when you ask:
synapse upgrade --with knowledge # an existing install
curl -LsSf …/setup.sh | bash -s -- --with knowledge # a new one
Then turn it on in Settings → General. Every later upgrade keeps it.
Upgrading
synapse upgrade
It backs up the database, installs the newest version, updates the database and starts Synapse again if it was running.
An npm install is upgraded with npm install -g synapse-orch-ai@latest; Docker by pulling the new image and running
the same command.
Coming from 1.x? Run the install command above, or synapse upgrade. Your agents, workflows, chats, schedules,
settings, API keys and vault files move into 2.0's store. Before anything changes, the upgrade backs up your 1.x data
folder, logs and .env to ~/.synapse/backups/synapse-1.x-<date>.tar.gz, with a manifest.json that lists where each
part came from; the old install's folder is removed once 2.0 is running. To go back, reinstall 1.9.3 and put each part
of the backup back where manifest.json says. Don't run 2.0's synapse uninstall first: it removes ~/.synapse,
backups included, unless you add --keep-data. Docker imports a 1.x /data volume on its first start. See
Upgrading to 2.0.
The command line
One synapse command runs SynapseOrch on this machine, and drives any SynapseOrch — this one, one on your servers, or
Cloud — with an API key. Everything you can do in the app, you can do from a terminal or a script.
This machine
synapse start # open it (-d: in the background)
synapse status # is it running? (--json for scripts)
synapse logs -f # follow the backend's log (frontend, install)
synapse upgrade # the newest version; data and settings stay
synapse autostart on # start when you log in
Build, run and answer
synapse agents list # list, show, create, update, delete
synapse workflows list # list, show, export, create, validate, deploy, delete
synapse run NAME -m "…" # run a workflow and follow it to its answer
synapse runs waiting # what waits on you: questions and held tool calls
synapse runs answer RUN_ID --approve # answer a sign-off (or --reject, or reply with text)
synapse chat --agent NAME -m "…" # one message to an agent (bare `synapse chat` to talk)
synapse schedules create --workflow NAME --cron "0 9 * * MON"
Knowledge, Boards and the workspace
synapse knowledge upload --base Handbook ./policies/*.pdf --wait
synapse knowledge search "refund window" --base Handbook
synapse boards show BOARD # a board's open cards; `boards card` adds and comments
synapse activity --days 7 # what ran, what failed, what waits
synapse usage --by model # what the workspace spent, by day, agent or model
Every door a key has
synapse tools list # every product tool this key may call; `tools call NAME`
synapse api GET /agents # call any route a key reaches, as `gh api` does
synapse mcp # serve SynapseOrch to Claude, Cursor… over stdio
synapse login --url https://api.synapseorch.com --api-key-stdin # drive another SynapseOrch
For scripts, --json prints JSON (one event a line for a run) and -q lists ids only. Exit codes say what happened:
0 done · 3 waiting on a person · 4 run failed · 5 not found · 6 key refused · 7 busy or unreachable.
synapse --help lists every command. CLI reference →
Use it from your own code and tools
- REST API — start runs, stream their events, answer sign-offs, with an API key. API →
- MCP server — build and run workflows from Claude, ChatGPT or Cursor. MCP server →
- OpenAI-compatible endpoint — every agent answers as a model to any OpenAI client, with its own tools.
Scale mode
When one process is not enough: a Redis job queue, 1 to 100+ workers, per-step Postgres checkpoints (a worker that dies loses nothing), S3-compatible storage and per-team run limits — behind the same API.
docker pull synapseorchai/synapse-ai-api-server:latest # stateless API server
docker pull synapseorchai/synapse-ai-worker:latest # workers, as many as needed
docker compose --profile scale up runs the whole stack (with REDIS_URL set in .env); Kubernetes examples are in
infra/k8s/.
Scale mode →
Running it on a network
Everything an agent can do runs on the machine Synapse runs on, so treat a shared instance like a server:
- turn on login (Settings → General) on anything reachable beyond your own machine;
- publish only the web port (
3000) on the network, never the backend's (8765); - local MCP servers and custom Python tools run commands: turn them off with
allow_stdio_mcpif you do not trust every user, and give agents the narrowest tools that do the job.
Security issues: please report them privately — see SECURITY.md.
SynapseOrch Cloud
The same product, run for you: no servers, managed models, plans with a hard monthly limit, and teams with roles. synapseorch.com
Contributing
See CONTRIBUTING.md for the development setup, how the pieces fit, and the PR checklist. Questions and ideas are welcome on Discord.
License
SynapseOrch is licensed under the GNU AGPL v3.