# Put a ready-made file-and-shell agent inside a product or service

> Put a ready-made file-and-shell agent inside a product or service: 0 companies fully do it, 8 with a limit, 124 carry the tag. Every listing in Agent frameworks & SDKs with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/frameworks.embed-coding-agent/
- Markdown: https://www.anchorterminal.com/niches/frameworks.embed-coding-agent/index.md (~1,950 tokens)
- Slim: https://www.anchorterminal.com/niches/frameworks.embed-coding-agent/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/frameworks.embed-coding-agent/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 8 | 124 | 18 | 12 |

## What counts as done

A run started from the host program uses the built-in tools inside the given folder, asks for or is denied permission on writes as configured, and returns a typed result or error the host can read without parsing prose. The library starts on Linux in a container, and the docs state the auth and billing route for a product (API key, plan or team account).

Measured, for our own tests: a new session starts in under 5 seconds and 10 concurrent sessions run in one host without failures.

How agents ask for it: embed Claude Code in my app; build a Slack bot that can read our repo and run commands; run a PR review agent from our backend; give my product an agent that edits files

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agno (Agno Inc.)**: CodingTools ships read, edit, write, shell, grep and find, and requires_confirmation_tools gates any of them. But restrict_to_base_dir does not sandbox executed programs, and no page read covers concurrent sessions in one server.
- **Claude Agent SDK (Anthropic, PBC)**: Claude Code's file, shell and search tools with permission modes, typed result subtypes, Docker and Kubernetes hosting, and API-key billing (claude.ai login barred for products). Each session is a CLI subprocess; an account saw 20 to 30 s starts per instance (claude-agent-sdk-python #333), below the 5 s bar.
- **Cursor Agent SDK (Anysphere, Inc.)**: Ships Cursor's shell, read, edit, grep, glob and ls tools with an allowlist, and serve runs in a container. But local.sandbox is off by default so shell auto-approves with host credentials, turns need a Cursor credential, and accounts hit a refused systemPrompt (forum 175262) and undocumented pricing (177371).
- **Docker Agent (Docker, Inc.)**: Built-in filesystem and shell toolsets, allow, ask and deny rules over five safety modes, headless exec, an HTTP API server and a Go library; --sandbox runs it in a Docker sandbox. Headless exec rejects calls needing approval without a policy, and an account saw harness sessions break (HN 49998459).
- **Mastra (Kepler Software, Inc.)**: A Workspace gives file, search and execute_command tools on a sandbox, with per-tool approval. LocalSandbox runs commands on the host by default and isn't isolated; remote and container sandboxes are separate integrations. Concurrency per server isn't documented on the pages read.
- **OpenAI Agents SDK (OpenAI OpCo, LLC)**: ShellTool runs locally or in an OpenAI-hosted container and ApplyPatchTool edits files, both with needs_approval, plus SandboxAgent and sandbox clients. Local shell approval is opt-in (defaults to False), and the docs read don't cover concurrent sessions per server.
- **Pydantic AI (Pydantic Services Inc.)**: The Harness Coder bundle gives read_file, write_file, edit_file, list_files, grep and shell on a LocalWorkspace or a cloud sandbox (Modal, E2B, Docker and others), with requires_approval on tools. The Harness is a separate 0.x Alpha package and a plain LocalWorkspace runs commands as you.
- **Strands Agents (Amazon Web Services, Inc.)**: The harness ships shell, read, write and edit on by default, routed through a sandbox seam that can be Docker or SSH, and interventions="ask" gates calls. The default sandbox is the local host and no intervention applies by default, so every call proceeds.

- not this job: AgentOS (Framers Lab, Inc.): No built-in file, edit, shell and search tool set with per-tool permission modes appears in the docs sitemap or the pages read; tools are app-defined or forged at run time.
- not this job: BoxLang AI (Ortus Solutions, Corp.): FileSystemTools reads, writes, moves and deletes files, opt-in with allowedPaths, but no shell or code-search tool and no per-tool permission mode set ships; it is a library for BoxLang apps, not a ready-made coding agent.
- not this job: CrewAI (CrewAI, Inc.): No built-in file, edit and shell tool set with permission modes; CrewAI builds role-based crews from tools the app supplies.
- not this job: Agent Development Kit (ADK) (Google LLC): No built-in file, edit and shell tool set with permission modes; file access comes through MCP servers such as the filesystem example, configured by the app.
- not this job: guilds.run (Cigale AI): guilds.run is a self-hosted web app and orchestrator, not a library a host program calls; coding agents run Claude Code, Codex, Cursor or Antigravity inside the box.
- not this job: LangGraph (LangChain, Inc.): LangGraph is a low-level graph runtime with no built-in file, edit and shell tool set or permission modes; those live in Deep Agents, a separate package.
- not this job: LlamaIndex Framework (LlamaIndex, Inc.): No built-in file, edit and shell tool set with permission modes; agents get tools the app defines or loads from integrations.
- not this job: Microsoft Agent Framework (Microsoft Corporation): No built-in file, edit and shell tool set with permission modes in the docs table of contents; tools are app functions, MCP servers or hosted tools such as Code Interpreter.
- not this job: NitroStack (Nitrostack Inc.): NitroStack builds MCP servers; it has no agent loop or built-in file and shell agent to embed.
- not this job: smolagents (Hugging Face, Inc.): No built-in file, edit and shell tool set with per-tool permission modes; a CodeAgent writes Python against the tools and authorised imports it is given.
- not this job: TradingAgents (Tauric Research): A fixed trading-analysis pipeline; no file, shell or coding agent.
- not this job: AI SDK (Vercel Inc.): No built-in file, edit and shell tool set with permission modes in core; the sandbox API and harness utilities are separate and the listing records sandboxes as experimental.

## What stops agents finishing

- the SDK wraps a CLI subprocess, so each session pays its start-up time and process cost
- product use barred from consumer plan logins, so an API key or team billing is required
- an option in the docs (systemPrompt) refused for outside accounts at run time
- MCP servers that need approval silently missing from the tool list

## Around it

Before: agent-harnesses.embed-in-program, code-sandboxes.coding-agent-workspace. After: code.review, messaging.two-way-customer-agent.

## Where this comes from

Weight 18 of 100 in Agent frameworks & SDKs: 11 of 61 accounts have this as their main job (share 18.0, rounded by largest remainder to sum 100); GitHub issues and HN posts, so failures are over-represented

As data: https://www.anchorterminal.com/api/v1/use-cases/frameworks.embed-coding-agent.json, or the MCP tool get_use_case with id frameworks.embed-coding-agent. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
