M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
O

Open Agent SDK (TypeScript)

Community
Framework / SDK
Open-source TypeScript Agent SDK that runs the full agent loop in-process, supporting Anthropic and OpenAI-compatible APIs.

This is a framework for building MCP servers or agents, not a server you connect a client to.

How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #412 of 438
Stars
★ 2.7k Very popular
Transport
stdio (local process)
Runtime
Node.js >=18
Credentials
API key / credential required
License
MIT
Last commit
Tools
35
30FMRS · D

Open Agent SDK (TypeScript) is a feature-rich open-source SDK offering an in-process agent loop, multi-model support, a wide range of tools, and a skill system. However, it is a third-party implementation, not officially maintained, and requires careful security configuration.

Strongest · Documentation 12/20 Weakest · Security and permissions 3/20

Reliability
4/20
Security and permissions
3/20
Maintenance
6/20
Documentation
12/20
Setup experience
5/20
Why each score
Reliability 4/20
Only the README and repository metadata are available; no source, CI, or tests can be verified. The README claims 35+ tools and a full agent loop, but there is no server manifest and no reproducible MCP initialization/tool-list handshake evidence. Error handling and dependency control cannot be confirmed. The static calibration caps this at 12, and evidence is insufficient, so scored 4.
Security and permissions 3/20
Red line triggered: the default permissionMode is bypassPermissions, and built-in tools include Bash, Write, and Edit, allowing irreversible delete/pay/deploy-style commands without confirmation. Passing the API key via environment variables is positive, but least privilege is absent by default, no sandbox is enabled by default, and external network and filesystem boundaries are unclear. Security scored 3.
Maintenance 6/20
The repository is not archived, has an MIT license, and 10 open issues, but no commit history, release records, or issue-response evidence is visible. Star count is only a discovery signal and earns no points. Lacking verifiable sustained maintenance and versioning evidence, scored 6.
Documentation 12/20
The README is fairly comprehensive: install steps, environment variables, API reference, 15 examples, architecture, and built-in tool tables are included. However, there is no troubleshooting section, no mainstream-client configuration example for using it as an MCP server, and no disclosure of the risks of the default bypassPermissions mode. It is good as a README, but as MCP-server documentation it has gaps; scored 12.
Setup experience 5/20
The SDK install steps are simple (npm install + environment variable), but there is no mainstream-client configuration example for adding this repository as an MCP server, nor is there a server manifest or executable-entry evidence. Examples in the README additionally require tsx. The static calibration allows above 15 only with verifiable execution evidence; here evidence is insufficient, so scored 5.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or codeUses the network

Best for

  • TypeScript projects that need an in-process agent loop
  • Developers wanting to support both Anthropic and OpenAI-compatible models
  • Applications requiring fine-grained control over tools and permissions
  • Agents to run in cloud, serverless, or Docker environments

Not for

  • Scenarios requiring official Claude Agent SDK maintenance
  • Users who need CLI functionality (SDK removes that dependency)
  • Users who want to use MCP servers with zero code

Required permissions

  • API key access (e.g., CODEANY_API_KEY)
  • File system read/write (depending on configured tools)
  • Network access (for API calls and WebFetch/WebSearch)
  • Environment variable configuration

Risks and side effects

  • Tools may perform destructive operations (e.g., Bash, Write), configure permissions carefully
  • API key leakage risk
  • Models may produce inaccurate or harmful outputs
  • Dependence on third-party API availability and costs

Setup

Before you start

Runtime：Node.js >=18

CODEANY_API_KEY requiredsecret API key required to call the LLM provider; obtain it from your API provider (Codeany, Anthropic-compatible, or OpenAI-compatible endpoint).
CODEANY_AUTH_TOKEN optionalsecret Optional alternative auth token that can be used instead of the API key.
Other optional settings (3)
CODEANY_API_TYPE optional Selects the API protocol: anthropic-messages (default) or openai-completions; auto-detected from model name if unset.
CODEANY_BASE_URL optional Custom API endpoint, e.g. https://openrouter.ai/api or https://api.openai.com/v1.
CODEANY_MODEL optional Default model override, e.g. gpt-4o or anthropic/claude-sonnet-4.
  1. Install the npm package: npm install @codeany/open-agent-sdk
  2. Set your API key environment variable: export CODEANY_API_KEY=your-api-key
  3. Configure other environment variables as needed (e.g., CODEANY_API_TYPE, CODEANY_BASE_URL, CODEANY_MODEL).
  4. Import the SDK in your code and use query or createAgent to get started.

Check that it works

After `npm install @codeany/open-agent-sdk` and setting CODEANY_API_KEY, run a simple `query()` or `createAgent().prompt()` call; a streamed model response confirms the SDK and provider connection work.

Troubleshooting

  1. Check API key is correctly set
  2. Verify API type and base URL configuration
  3. Ensure Node.js version >= 18
  4. Check if model name supports auto-detection
  5. Inspect network connectivity and proxy settings
  6. Check disk space for session persistence

Things to try

Once connected, you can ask your AI assistant things like:

  • Read package. and tell me the project name.
  • Create a file /tmp/hello.txt with "Hello World", then read it back.
  • What is the weather in Tokyo?
  • Use the code-reviewer agent to review src/index.ts.

Tools 35

Bash writes
Execute shell commands
Read read-only
Read files with line numbers
Write writes
Create or overwrite files
Edit writes
Precise string replacement in files
Glob read-only
Find files by pattern
Grep read-only
Search file contents with regex
WebFetch read-only
Fetch and parse web content
WebSearch read-only
Search the web
Show 27 more tools
NotebookEdit writes
Edit Jupyter notebook cells
Agent writes
Spawn subagents for parallel work
Skill writes
Invoke registered skills
TaskCreate writes
Create a task
TaskList read-only
List tasks
TaskUpdate writes
Update a task
TaskGet read-only
Get a task
TaskStop writes
Stop a task
TaskOutput read-only
Get task output
TeamCreate writes
Create a multi-agent team
TeamDelete destructive
Delete a multi-agent team
SendMessage writes
Send messages between agents
EnterWorktree writes
Enter a Git worktree isolation environment
ExitWorktree writes
Exit a Git worktree isolation environment
EnterPlanMode writes
Enter structured planning mode
ExitPlanMode writes
Exit structured planning mode
AskUserQuestion read-only
Ask the user for input
ToolSearch read-only
Discover lazy-loaded tools
ListMcpResources read-only
List MCP resources
ReadMcpResource read-only
Read MCP resources
CronCreate writes
Create a scheduled task
CronDelete destructive
Delete a scheduled task
CronList read-only
List scheduled tasks
RemoteTrigger writes
Remote agent triggers
LSP read-only
Language Server Protocol (code intelligence)
Config writes
Dynamic configuration
TodoWrite writes
Session todo list

Use cases

Building autonomous agent applications
Executing multi-step tool-calling tasks
Integrating MCP servers
Developing custom skills and tools
Implementing chat interfaces or automation workflows

Supported clients

Claude Desktop

Listed from the project's documentation, not tested by this site.

Overview

Open Agent SDK (TypeScript) is an open-source SDK that lets developers run a full agent loop in their own application without needing a CLI subprocess. It supports the Anthropic Messages API and OpenAI-compatible APIs (such as OpenAI, DeepSeek, Qwen, Mistral), and offers a rich set of built-in tools, skill system, lifecycle hooks, subagents, permissions, session persistence, and more. It can be deployed in cloud, serverless, Docker, and CI/CD environments.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.3k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision a9ade5fdaefb Data synced 2026-10-19 Read the FMRS scoring method