Built-in Tools
Tools are functions that agents can call during their ReAct loop.
Tools are functions an agent can call during its ReAct loop. SynapseOrch ships a set that needs no setup, plus your own and any MCP server you connect.
Available tools
| Tool | Description | Requires |
|---|---|---|
| Sandbox | Execute Python in an isolated Docker container | Docker |
| Vault | Read and write persistent files | — |
| SQL | Query any SQLAlchemy-compatible database | DB config |
| Browser | A browser inside Synapse that agents work in while you watch | Node, and Playwright's Chromium |
| Web Scraper | Anti-bot evasion web scraping | — |
| Code Search | Semantic search over code repositories | Postgres + PG Vector + repo indexed |
| PDF Parser | Extract text and tables from PDF URLs | — |
| Excel Parser | Parse .xlsx files | — |
| Collect Data | Generate dynamic input forms | — |
| Time | Natural language date/time parsing | — |
Always available
Every agent also carries a few tools that are part of the engine rather than optional capabilities:
| Tool | Does |
|---|---|
update_tasks | The task checklist you see fill in during a run |
load_skill | Loads one of the agent's skills on demand |
remember · read_memory · search_memory · forget_memory | Only on agents with memory enabled |
spawn_subtask | Helpers: short-lived agents for one task, when Can send helpers is on |
How tools work
Tools reach agents three different ways, and the difference matters when something fails to start:
- In-process — the sandbox, SQL, the parsers, the scraper, code search and the rest run inside the backend. Nothing to install, nothing to keep alive.
- Engine-native —
update_tasks,load_skill,spawn_subtaskand the memory tools are imported directly by the engine. Deliberately not MCP servers, so there is no subprocess and they survive a resume. - Automatically started MCP servers — Filesystem and Google Workspace (once it has credentials) are launched at startup. These are real subprocesses and can fail to start independently of the backend. There is no separate memory server: agent memory is built into the engine.
- The browser — Browser Automation (while browser automation is on, as it is by default) starts on an agent's first browser action, not at startup: its own process, with one Chromium for each chat or run. It needs Node and Playwright's Chromium on the machine. See Browser.
External MCP servers you connect yourself work exactly like that last group — see MCP servers.
Assigning tools to agents
In the agent editor, set tools to:
["all"]— enable all available tools- A named list — e.g.
["vault_write", "run_sql_query", "parse_pdf"]
In workflow steps, use allowed_tools to further narrow an agent's access for that specific step:
"allowed_tools": ["vault_read", "vault_write"]Tool names
These are the names to use when giving an agent an explicit tool list, or when naming a tool in a prompt. They do not follow a single prefix convention — use the name as written.
| Group | Tools |
|---|---|
| Sandbox | execute_python · tool_code |
| Vault (write) | vault_read · vault_write · vault_patch · vault_create · vault_list · vault_delete |
| Vault (read) | read_file · read_json · list_files · search_files |
| SQL | run_sql_query · list_tables · get_table_schema |
| Web Scraper | scrape_url · scrape_structured · scrape_with_session · crawl_multiple · extract_links · search_page · screenshot_url |
| Code Search | search_codebase · grep · glob · find_similar_code · get_file_chunks · read_file_by_lines · multi_repo_search · list_indexed_files |
| PDF Parser | parse_pdf |
| Excel Parser | parse_xlsx |
| Collect Data | collect_data |
| Time | get_datetime |
| Bash | run_bash |
| Personal details | get_personal_details |
The Browser tools are listed on Browser.
Bash
run_bash runs a shell command in one of the folders agents may work in — linked repositories, the vault and Settings → General → Allowed directories — starting in the first repository, or the vault, when the call names no folder. Commands from any chat, step or branch run at the same time. Each has a timeout, 30 seconds by default and at most 120, and a command that runs past it is killed together with everything it started — so detach a background server with > log 2>&1 &. A command's stdin is empty. Stopping a chat turn does not stop a command already running: it runs to its timeout.
To see everything available, including external MCP tools, open Apps › Tools.
Custom tools
Beyond native tools, you can register REST API tools and Python tools as first-class agent tools. External MCP servers also expose their tools to agents automatically.
Boards and Agents
Which boards an agent reads and works, what it sees at the start of each turn, how it claims and closes tasks, what a person's card means to it, and the From content mark.
Python Sandbox
The sandbox tool executes Python code in an isolated Docker container with resource limits. It's the safest way to give agents the ability to run code without exposing your host system.