gcontext

The framework for building stateful agents.

An agent built with gcontext is a folder: instructions, service connections, secrets, knowledge, and multi-step work, all as plain files you can version with git. gcontext serves that folder over MCP from a local HTTP server, and you use the agent from the tools you already work in: Claude Code, Claude Desktop, Codex, or Cursor.

Runtimes forget everything between sessions; the folder doesn't. Because the state is separate from the runtime, the same agent works from any client and survives every session. gcontext ships no chat loop and no LLM client: the runtime you attach does the reasoning, gcontext keeps the state.

Install

uv tool install gcontext-ai

Quickstart

gcontext init my-agent      # create the state folder
gcontext up my-agent        # serve it at http://127.0.0.1:4242/mcp

Then connect a client (once, from any directory):

claude mcp add --transport http my-agent http://127.0.0.1:4242/mcp

gcontext connect claude|desktop|codex|cursor prints the exact steps per client. The server logs each client as it connects. Stopping the server (Ctrl+C) disconnects everything; there is no other cleanup.

The folder

my-agent/
  gcontext.yaml          # name, description, optional port
  instructions.md        # pushed to every agent at connect: what it starts with
  secrets.env            # secret values, gitignored

  connections/           # services the agent can use
    stripe/
      connection.yaml    # secret names + Python deps
      index.md           # API notes, usage patterns

  modules/               # accumulated knowledge
  archive/               # excluded from scanning, still readable

Markdown holds the context, YAML holds the config. Edit any of it with a text editor; the server reads the files on demand, so changes apply immediately. Two exceptions load at server start and need a restart to pick up edits: instructions.md (pushed in the MCP handshake) and command files.

Connected clients get five tools: read_file, write_file, list_dir, grep, run_script.

run_script runs either ad-hoc code or a saved script by path (scripts/ folders hold proven procedures, so they are reused instead of rewritten). Files under connections/*/commands/ and modules/*/commands/ register as MCP prompts, which Claude Code shows as slash commands; see "Commands" below.

Your first connection

init creates no connections: a connection is worth having when it points at a service you actually use. Adding one is three files, no command needed:

mkdir -p my-agent/connections/stripe

connections/stripe/connection.yaml declares what the connection needs, by name only:

name: stripe
description: Payments, test mode.
secrets:
  - STRIPE_API_KEY
deps:
  - stripe

Put the value in secrets.env (gitignored, never leaves your machine):

echo 'STRIPE_API_KEY=sk_test_...' >> my-agent/secrets.env

And write connections/stripe/index.md: what the service is for, which endpoints matter, any usage patterns worth remembering. The agent reads this before writing scripts, and updates it as it learns.

That's it. The server picks the connection up on the next tool call (no restart), gcontext status shows whether every declared secret has a value, and the agent can now call the API through run_script without ever seeing the key.

Context ledger

gcontext context lists every channel through which context reaches the agent, marked as loaded (pushed at connect), on demand (agent pulls it via a visible tool call), skipped (nothing to push), or uncontrolled (owned by the runtime, outside gcontext's view). gcontext only inserts context through the channels on that list. If you want to know what the agent is seeing, this is the answer.

Controlled session

The ledger marks runtime-owned pipes (the runtime's system prompt, its config files, its other MCP servers) as uncontrolled, because gcontext cannot close them. If you want a claude session with those pipes closed, launch claude yourself with its own flags; there is no gcontext command for this, since it is a runtime invocation, not framework behavior:

claude --mcp-config '{"mcpServers":{"gcontext":{"type":"http","url":"http://127.0.0.1:4242/mcp"}}}' \
       --strict-mcp-config \
       --setting-sources ""

--strict-mcp-config ignores every other configured MCP server, and --setting-sources "" skips CLAUDE.md files and user settings. Your instructions.md still arrives through the MCP handshake, like in any session. Adjust the URL to your project's port.

Secrets

connection.yaml declares secret names; secrets.env holds the values. When the agent calls run_script, the values are injected as environment variables and scrubbed from the script's output. The agent can know that STRIPE_API_KEY exists and use it in a script, but never reads the value. secrets.env is gitignored by init and the write_file tool refuses to touch it.

run_script executes Python in a per-project venv with each connection's declared deps preinstalled (via uv).

Archiving

When old modules or connections start cluttering the context, move them:

mv my-agent/modules/old-onboarding my-agent/archive/modules/

Anything under archive/ is skipped when scanning, but stays readable by path, and summaries mention what's archived so it doesn't silently vanish. That's the entire mechanism. gcontext never moves, archives, or deletes anything on its own.

Commands

A command is a user-invokable entry point stored next to the knowledge it belongs to: a file under connections/<name>/commands/ or modules/<name>/commands/. The server registers each one as an MCP prompt named <owner>__<command>; Claude Code shows it as a slash command (/mcp__<server>__<owner>__<command>). Prompts cost no tool-schema context: a command's text enters the conversation only when you invoke it.

Two file types:

• .md: YAML frontmatter (description, parameters), then the body that gets injected, with $name placeholders filled from the arguments.

  ---
  description: Draft a refund reply
  parameters:
    - name: email
      required: true
  ---
  Draft a refund reply for $email and show it to the user.
  

• .py: a runnable script with the same frontmatter as a # --- comment block at the top. Invoking it instructs the agent to run the file through run_script, with the arguments passed as params (they reach the script as PARAM_<NAME> env vars).

Commands are discovered at server start; restart to pick up new files.

Dashboard

gcontext up also serves a read-only dashboard at the server root, for example http://127.0.0.1:4242/. It shows the project overview and context ledger, connections with secret status (names only, never values), modules, commands, a file browser, and a live activity feed of every tool call agents make. The feed lives in server memory and empties on restart. The dashboard changes nothing; agents make the changes.

Developing the dashboard itself needs node: make web-dev runs a Vite dev server on http://localhost:5179 that proxies to the gcontext server, and make web-build produces the static bundle that gcontext up serves.

CLI

Command	Description
gcontext init <dir>	Scaffold a new state folder
gcontext up [dir]	Serve the folder over MCP
gcontext status [dir]	Server state, connected clients, state overview
gcontext connect [client]	Connection steps for claude, desktop, codex, cursor
gcontext context [dir]	Print the context ledger

Going further

• examples/ops-agent: a complete agent folder with connections, modules, a command, and an archived module
• docs/design.md: why gcontext is built this way, decision by decision
• docs/modules.md: writing portable, shareable modules

Scope

Local only. The server binds 127.0.0.1 without auth, so it is not reachable from outside your machine and should stay that way. A remote variant (same model, URL plus token) is planned but not part of this release.

License

MIT