@devries-cameron20 / composio-1 / commits / 6d2393316
Fork 0

Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.

docs: add Workbench concept page

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Sushmithamallesh committed Feb 11, 2026 at 18:55 UTC 6d23933167517f1bcf7c62a04652920d6621287d
4 files changed +90 -5
docs/content/docs/meta.json
+1
@@ -8,6 +8,7 @@
8 "users-and-sessions",
9 "authentication",
10 "tools-and-toolkits",
11 + "workbench",
12 "triggers",
13 "---Getting Started---",
14 "configuring-sessions",
docs/content/docs/tools-and-toolkits.mdx
+2 -2
@@ -15,7 +15,7 @@ When you create a session, your agent gets these 5 meta tools:
15 | `COMPOSIO_SEARCH_TOOLS` | Discover relevant tools across 800+ apps |
16 | `COMPOSIO_MANAGE_CONNECTIONS` | Handle OAuth and API key authentication |
17 | `COMPOSIO_MULTI_EXECUTE_TOOL` | Execute up to 20 tools in parallel |
18 -| `COMPOSIO_REMOTE_WORKBENCH` | Run Python code in a persistent sandbox |
18 +| `COMPOSIO_REMOTE_WORKBENCH` | Run Python code in a [persistent sandbox](/docs/workbench) |
19 | `COMPOSIO_REMOTE_BASH_TOOL` | Execute bash commands for file and data processing |
20
21 Meta tool calls in a session are correlated using a `session_id`, allowing them to share context. The tools can also store useful information (like IDs and relationships discovered during execution) in memory for subsequent calls.
@@ -54,7 +54,7 @@ Done. (For large results, agent can use REMOTE_WORKBENCH to process)
54
55 For most tasks, `COMPOSIO_MULTI_EXECUTE_TOOL` returns results directly. But when dealing with large responses or bulk operations, your agent uses the workbench tools:
56
57 -- **`COMPOSIO_REMOTE_WORKBENCH`** - Run Python code in a persistent sandbox. Use for bulk operations (e.g., labeling 100 emails), complex data transformations, or when results need further analysis with helper functions like `invoke_llm`.
57 +- **`COMPOSIO_REMOTE_WORKBENCH`** - Run Python code in a [persistent sandbox](/docs/workbench). Use for bulk operations (e.g., labeling 100 emails), complex data transformations, or when results need further analysis with helper functions like `invoke_llm`.
58
59 - **`COMPOSIO_REMOTE_BASH_TOOL`** - Execute bash commands for simpler file operations and data extraction using tools like `jq`, `awk`, `sed`, and `grep`.
60
docs/content/docs/users-and-sessions.mdx
+4 -3
@@ -97,6 +97,7 @@ const toolkits = await session.toolkits();
97 </Tab>
98 </Tabs>
99
100 -For details on enabling toolkits, setting auth configs, and using session methods:
101 -
102 -<Card icon={<BookOpen />} title="Configuring Sessions" href="/docs/configuring-sessions" description="Enable toolkits, set auth configs, and select connected accounts" />
100 +<Cards>
101 + <Card icon={<BookOpen />} title="Configuring Sessions" href="/docs/configuring-sessions" description="Enable toolkits, set auth configs, and select connected accounts" />
102 + <Card icon={<BookOpen />} title="Workbench" href="/docs/workbench" description="Write and run code in a persistent sandbox" />
103 +</Cards>
docs/content/docs/workbench.mdx new
+83
@@ -0,0 +1,83 @@
1 +---
2 +title: Workbench
3 +description: A persistent Python sandbox for writing code, calling tools programmatically, and processing data
4 +keywords: [workbench, sandbox, remote execution, bulk operations, data processing]
5 +---
6 +
7 +The workbench is a persistent Python sandbox where your agent can write and execute code. It has access to all Composio tools programmatically, plus helper functions for calling LLMs, uploading files, and making API requests. State persists across calls within a session. The `COMPOSIO_REMOTE_BASH_TOOL` meta tool also runs commands in the same sandbox.
8 +
9 +<Callout>
10 +The workbench is part of the meta tools system. It's available when you create sessions, not when [executing tools directly](/docs/tools-direct/executing-tools).
11 +</Callout>
12 +
13 +## Where it fits
14 +
15 +Your agent starts with `SEARCH_TOOLS` to find the right tools, then uses `MULTI_EXECUTE` for straightforward calls. When the task involves bulk operations, data transformations, or multi-step logic, the agent uses `COMPOSIO_REMOTE_WORKBENCH` instead.
16 +
17 +```mermaid
18 +graph TD
19 + U["Triage my emails, label urgent ones, log to a Google Sheet"] --> S["1. SEARCH_TOOLS"]
20 + S --> M["2. MULTI_EXECUTE"]
21 + M --> W["3. REMOTE_WORKBENCH"]
22 + W --> D["Done"]
23 +
24 + S -.- S1["Discovers Gmail and Sheets tools"]:::annotation
25 + M -.- M1["Fetches unread emails"]:::annotation
26 + W -.- W1["Classifies, labels, and logs to sheet in parallel"]:::annotation
27 +
28 + classDef annotation stroke-dasharray: 5 5
29 +```
30 +
31 +## What the sandbox provides
32 +
33 +### Built-in helpers
34 +
35 +These functions are pre-initialized in every sandbox:
36 +
37 +| Helper | What it does |
38 +|--------|-------------|
39 +| `run_composio_tool` | Execute any Composio tool (e.g., `GMAIL_SEND_EMAIL`, `SLACK_SEND_MESSAGE`) and get structured results |
40 +| `invoke_llm` | Call an LLM for classification, summarization, content generation, or data extraction |
41 +| `upload_local_file` | Upload generated files (reports, CSVs, images) to cloud storage and get a download URL |
42 +| `proxy_execute` | Make direct API calls to connected services when no pre-built tool exists |
43 +| `web_search` | Search the web and return results for research or data enrichment |
44 +| `smart_file_extract` | Extract text from PDFs, images, and other file formats in the sandbox |
45 +
46 +### Pre-installed libraries
47 +
48 +The sandbox has these libraries pre-installed:
49 +
50 +- **Data analysis**: pandas, numpy, matplotlib
51 +- **Image processing**: Pillow, OpenCV, scikit-image
52 +- **Machine learning**: PyTorch
53 +- **Document handling**: python-docx, pdfplumber, reportlab, pandoc
54 +- **Standard tools**: requests, json, csv, and the Python standard library
55 +
56 +### Persistent state
57 +
58 +The sandbox runs as a persistent Jupyter notebook. Variables, imports, files, and in-memory state from one call are available in the next.
59 +
60 +## Common patterns
61 +
62 +### Bulk operations across apps
63 +
64 +Some tasks touch hundreds of items across services. Say you need to triage 150 unread emails. The agent writes a workbench script: classify each email with `invoke_llm`, apply Gmail labels with `run_composio_tool`, and log results to a Google sheet.
65 +
66 +### Data analysis and reporting
67 +
68 +The agent can chain tools inside the sandbox. Fetch GitHub activity, aggregate with pandas, chart with matplotlib, summarize with `invoke_llm`, upload a PDF with `upload_local_file`.
69 +
70 +### Multi-step workflows
71 +
72 +The sandbox preserves variables and files across calls. The agent can paginate through records, transform them, and write to a destination over multiple calls.
73 +
74 +## Related
75 +
76 +<Cards>
77 + <Card icon={<Blocks />} title="Tools and toolkits" href="/docs/tools-and-toolkits">
78 + How meta tools discover, authenticate, and execute tools at runtime
79 + </Card>
80 + <Card icon={<BookOpen />} title="Browse toolkits" href="/toolkits">
81 + Explore all available toolkits
82 + </Card>
83 +</Cards>