Claude Projects API: What You Can Access

Claude Projects is a feature in the Claude.ai web interface that lets you group conversations, upload files, and add persistent instructions that apply to every chat in the project. It is a product feature, not a separate API endpoint. When you call the Anthropic API directly, you do not get Projects; you get a conversation API where you manage system prompts and message history yourself.

That distinction matters if you are building an application or workflow on top of Claude. The API gives you the underlying model capabilities, and you reconstruct the project-like behavior — persistent instructions, shared files, saved context — in your own code or through an external record. Tools like fenbs fill part of that gap: they keep the project record outside the model and serve it to connected assistants through MCP, so the context follows the work rather than living in a web UI that the API cannot reach.

What Claude Projects are

In Claude.ai, a Project is a container with three parts: a set of conversations, a system prompt you write once and that applies to every new chat, and files you upload that Claude can reference. When you start a new conversation inside a project, Claude already has the instructions and can search the uploaded files.

On September 17, 2026, Claude.ai Projects added shared memory: team members on a shared plan can create a project where each person's conversations draw from the same instructions and file set. Each conversation still has its own chat history, but the project-level content is shared.

For more on what Projects are, see that guide.

There is no Projects API

The Anthropic API exposes the language model and its tools: you send messages, receive completions, and optionally pass tool definitions. It does not expose your Projects, the files you uploaded to them, or the custom instructions you wrote in the Claude.ai interface. Those are UI-layer features that do not have corresponding API endpoints.

This means an application you build on the Anthropic API cannot read your Claude.ai Projects or inherit their instructions. If you want a Claude.ai Project and an API-based application to share the same instructions, you maintain them in both places separately.

The Claude API guide covers what the API does expose: models, message structure, tools, and the system prompt. The API key guide covers authentication.

Replicate project instructions through the system prompt

The system prompt is the API's equivalent of the project instructions field. Text you put in the system prompt applies to every message in that conversation. For a persistent project context, store the instructions in a file or a database your application reads at the start of each session, then inject them into the system prompt.

Keep the instructions short enough to stay within a reasonable portion of the context window. Long, detailed instructions are harder to maintain and harder for the model to follow precisely. Write the rules, not the reasoning, and keep the source material in files rather than in the prompt.

Files in the API: the Files API and context uploads

Anthropic provides a Files API that lets you upload a file once and reference it in multiple API calls by ID, rather than resending the bytes every time. This is the API-side equivalent of uploading a document to a Claude.ai Project. You upload the file, save the file ID, and include it in the content of a user or assistant message.

The Files API supports PDFs, text files, and several other formats. Files are not searchable by keyword out of the box — Claude reads them as part of the context you include in a call. For retrieval-augmented workflows where you need to search a large document collection, you would implement a retrieval layer yourself or use a purpose-built service. Check Anthropic's documentation↗ (opens in a new tab) for current format support and limits.

Claude Code: AGENTS.md and shared project memory

Claude Code, Anthropic's agentic coding assistant, handles project context differently from the web UI. Claude Code reads an AGENTS.md file in your repository natively — it looks for that file at the start of each session and treats its contents as instructions. This is the Claude Code equivalent of project instructions, and it is version-controlled alongside your code.

On September 17, 2026, Claude Code Projects added shared memory for teams: you can create a Claude Code Project where multiple engineers share the same base instructions. Each engineer's sessions still run independently, but the project-level AGENTS.md is shared. This is separate from the Projects feature in Claude.ai.

For Claude Code-specific context, see the Claude Code documentation guide and the Claude project memory guide.

Manage the project record across sessions

The main challenge with any project context that lives outside the model is keeping it current. A system prompt or an AGENTS.md file that was accurate three months ago may not reflect the current rules, the current task list, or the current state of the work. Stale context is worse than no context because the assistant follows the wrong instructions confidently.

Build a habit of updating the project record when a decision changes. If you are using a file-based approach, version-control the instructions and review them when starting a new phase. If you are using an external record, update it as part of finishing a task so the next session starts from the current state.

For managing project context at the API level in an application you are building, see the Claude for project management guide. For Projects inside Claude.ai, the Claude Projects in cowork guide covers the shared setup.

How to do this with fenbs

1. Create a free account and add a project that mirrors the project you are building or managing.
2. In Decisions and rules, record the instructions that belong in your system prompt or AGENTS.md — the rules that apply to every session, written once and kept current.
3. Connect Claude or Claude Code over MCP from Integrations so each session reads the project record before starting, rather than relying on a static file.
4. Add context notes for the external files, repositories, and links the project depends on, so the assistant can find them when it needs them.
5. Before each session ends, ask the assistant to save progress on the tasks it touched and update Where we left off so the next session starts from the current state.

fenbs does not replace the Anthropic API or extend it. It provides the project record that a connected assistant reads through MCP. For the API itself, see the Claude API guide and Anthropic's developer documentation↗ (opens in a new tab).