AI-Assisted Development

Use AI coding tools to build Bifrost workflows, forms, and apps

There are five ways to use AI tools with Bifrost. Pick whichever fits — they all talk to the same platform.

Approach	Best for	Requirements
Claude Code + Bifrost plugin	Developers with local repos and git workflows	Claude Code, Bifrost SDK installed
Codex + Bifrost plugin	Codex users who want the same public Bifrost skills as Claude Code	Codex, Bifrost SDK installed
Other coding agents + Bifrost skills	Copilot CLI, Cursor, OpenCode, Gemini CLI, and other tools that read portable agent skills	Bifrost CLI, Python 3.11+
Local SDK	Any AI tool (Copilot, Cursor, etc.) with local files	Bifrost SDK, Python 3.11+
MCP (Remote)	Non-developers, quick edits, no local setup	MCP-compatible AI tool

1. Claude Code with the Bifrost plugin

The fastest path if you use Claude Code. The Bifrost plugin installs the public Bifrost skills:

/bifrost:build — create and debug workflows, forms, agents, and apps with the CLI and MCP.
/bifrost:setup — install the CLI, authenticate, and configure MCP on a fresh machine.
/bifrost:copilot-cowork-package — package a Bifrost agent as a Microsoft 365 Copilot Cowork plugin.
/bifrost:migrate — move a legacy v1 inline app and its backing entities into an installable v2 Solution.

Install the plugin

The recommended path is the Claude Code plugin marketplace. It auto-updates when you refresh the marketplace, so you stay in lockstep with the latest skill content without re-running install commands.

Inside Claude Code, run:

/plugin marketplace add https://github.com/gobifrost/bifrost
/plugin install bifrost@bifrost

To pull the latest content later, run /plugin marketplace update bifrost — or enable auto-update in /plugin → Marketplaces.

2. Codex with the Bifrost plugin

Codex can install the same public Bifrost plugin from GitHub. This gives Codex the same public skill set as Claude Code: build, setup, copilot-cowork-package, and migrate.

Inside Codex, run:

codex plugin marketplace add gobifrost/bifrost --ref main
codex plugin add bifrost@bifrost

To refresh a Git-backed marketplace later:

codex plugin marketplace upgrade bifrost
codex plugin add bifrost@bifrost

3. Other Coding Agents with `bifrost skill`

For non-Claude-Code and non-Codex harnesses, install via the Bifrost CLI. It pulls the public skills from GitHub (no clone required) and writes them to both <cwd>/.claude/skills/ and <cwd>/.agents/skills/ — the cross-harness portable convention read by Copilot CLI, Cursor, OpenCode, Gemini CLI, and other tools:

bifrost skill list                   # show what's currently installed
bifrost skill update                 # install or refresh all public skills
bifrost skill update --ref v1.4.2    # pin to a specific release
bifrost skill remove bifrost-build   # uninstall a skill from both locations

bifrost skill update only installs the publicly-distributed skills (bifrost-build, bifrost-setup, bifrost-copilot-cowork-package, bifrost-migrate) — internal maintainer skills stay in the repo. Re-run it whenever you want the latest content; there’s no auto-update, by design.

Migrating from a hand-copied skill

Earlier versions of the bifrost:build skill were distributed by cloning the bifrost repo and copying .claude/skills/bifrost-build/ into your project. If you set things up that way, remove the local copy before installing via the plugin or bifrost skill update so the two don’t drift:

rm -rf .claude/skills/bifrost-build .claude/skills/bifrost-setup .claude/skills/bifrost-copilot-cowork-package .claude/skills/bifrost-migrate
rm -rf .agents/skills/bifrost-build .agents/skills/bifrost-setup .agents/skills/bifrost-copilot-cowork-package .agents/skills/bifrost-migrate

Then run the plugin install or bifrost skill update and you’re current.

Setup after installing a plugin or skill

After the plugin or skill is installed, wire up the CLI and (optionally) the MCP server:

# Install the SDK from your instance and authenticate
pipx install https://your-instance.gobifrost.com/api/cli/download/bifrost-cli.tar.gz
bifrost login --url https://your-instance.gobifrost.com

# (Optional) Add the MCP server to Claude Code for chat-augmented exploration of platform state
claude mcp add --transport http bifrost https://your-instance.gobifrost.com/mcp

Or run the setup skill in your coding agent and it walks you through everything. In Claude Code, that is /bifrost:setup.

How it works

The build skill gives coding agents two modes:

SDK-first (local development):

Start bifrost watch to auto-push file changes
Write workflow code locally in your git repo
Files auto-push as you save — no manual sync needed
Test with bifrost run <file> -w <function> -p '{...}'
Use the entity CLI commands (bifrost forms create, bifrost agents update, bifrost configs set, …) or bifrost api for raw API calls
Commit and push to GitHub when ready

MCP-only (remote development):

Call bifrost_find_agents with the task you need to perform
Load the selected agent with bifrost_get_agent
Inspect the chosen tool with bifrost_get_tool_schema
Run it with bifrost_execute_tool and iterate through the same agent boundary

The skill checks integrations, reads SDK documentation, validates before declaring ready, and asks about org scoping.

CLI commands for platform operations

bifrost <entity> covers entity mutations (orgs, roles, workflows, forms, agents, apps, configs, integrations, events, tables). bifrost api is the escape hatch for any endpoint not yet wrapped by a typed command:

# Download SDK documentation (once per session, then grep locally)
bifrost api GET /api/docs/sdk > /tmp/bifrost-docs/sdk.md

# Execute a workflow synchronously
bifrost api POST /api/workflows/{id}/execute '{"workflow_id":"...","input_data":{...},"sync":true}'

# Check an execution
bifrost api GET /api/executions/{id}

# List platform state (debug sync divergence)
bifrost workflows list
bifrost api GET /api/workflows

For the full CLI command index, see the CLI cheatsheet.

When to use which

Surface	Use it for
CLI (`bifrost <entity> ...`, `bifrost run`, `bifrost watch`)	All entity mutations and file sync. First choice for repeatable, scriptable work.
MCP gateway (`bifrost_find_agents`, `bifrost_get_agent`, `bifrost_get_tool_schema`, `bifrost_execute_tool`)	Chat-augmented discovery and execution through the agents and tools your role can access.
`bifrost api`	Endpoints with no dedicated CLI command yet (e.g. arbitrary admin endpoints, debug routes, analytics queries).

Forms, apps, and agents are first-class CLI citizens (bifrost forms create, bifrost apps create, bifrost agents create) — they no longer require MCP. Use whichever surface fits the task.

4. Local SDK Development

Works with any AI tool that can edit local files (GitHub Copilot, Cursor, Windsurf, etc.).

Setup

pipx install https://your-instance.gobifrost.com/api/cli/download/bifrost-cli.tar.gz
bifrost login --url https://your-instance.gobifrost.com

The login command opens your browser for authentication. Credentials refresh automatically and use the system credential manager when available, with a platform-specific file fallback.

Workflow

Write Python files with @workflow, @tool, or @data_provider decorators
Start bifrost watch in your workspace to auto-push changes
Register newly-decorated functions with bifrost workflows register --path <file> --function-name <name>
Test locally: bifrost run my_workflow.py -w hello_world -p '{"name": "Alice"}'
All SDK modules (ai, integrations, config, knowledge, etc.) work locally — they call the remote API
Commit and push to git when ready

What to tell your AI tool

Give your AI assistant this context:

I'm building workflows for Bifrost, a Python automation platform.
- Workflows use `@workflow`, `@tool`, or `@data_provider` decorators from the `bifrost` package
- All functions must be async
- SDK modules: bifrost.ai, bifrost.config, bifrost.integrations, bifrost.knowledge, bifrost.tables, bifrost.files, bifrost.users, bifrost.organizations, bifrost.roles, bifrost.executions, bifrost.forms, bifrost.workflows
- Use `from bifrost import context` to access context.org_id, context.user_id, context.email
- Use `logging.getLogger(__name__)` for execution logs
- Return dicts or Pydantic models
- Entity mutations (orgs, roles, forms, agents, apps, configs, integrations, events, tables) all have first-class CLI commands: `bifrost <entity> <verb>`. See https://docs.gobifrost.com/sdk-reference/cli/ for the full surface.

5. MCP for External AI Tools

Connect Claude Desktop, ChatGPT, or any MCP-compatible tool directly to Bifrost. No local files needed.

Setup

Add the Bifrost MCP server to your AI tool’s configuration. For Claude Desktop, add to your config file:

{
  "mcpServers": {
    "bifrost": {
      "type": "http",
      "url": "https://your-instance.gobifrost.com/mcp"
    }
  }
}

You can also enable MCP from within Bifrost at Settings → MCP.

What MCP gives you

The root endpoint gives every client four stable tools:

Discover: bifrost_find_agents finds relevant agents that the signed-in user can access.
Load: bifrost_get_agent returns the selected agent’s live instructions and permission-filtered tool catalog.
Inspect: bifrost_get_tool_schema returns the current input schema for one tool.
Execute: bifrost_execute_tool dispatches the tool through that agent and returns an auditable result envelope.

This catalog does not change whenever a workflow or system tool is added. If you need one agent’s native tools to appear directly in the client, connect to https://your-instance.gobifrost.com/mcp/{agent_id} instead.

MCP system prompt

Copy this into your AI tool’s system instructions for best results:

You help build automations on the Bifrost platform through its MCP agent gateway.

1. Call `bifrost_find_agents` with the user's complete task.
2. Select the most relevant accessible agent and call `bifrost_get_agent`.
3. Follow that agent's live instructions. Do not invent tools that are not in its catalog.
4. Before each tool call, use `bifrost_get_tool_schema` and supply only schema-valid arguments.
5. Execute with `bifrost_execute_tool` using the same agent id and tool reference.
6. If the gateway returns repair guidance, correct the request instead of bypassing the agent boundary.

For creation or mutation work, confirm organization scope before executing the write. Publish apps only when the user explicitly asks.

Choosing an Approach

Use Claude Code or Codex + the Bifrost plugin for the best experience — the public skills handle mode switching, validation, setup, and testing automatically.

Use Other Coding Agents + bifrost skill update when your agent supports portable .agents/skills but not the Claude or Codex plugin marketplace.

Use Local SDK if you prefer a different AI tool or want full control over your git workflow.

Use MCP if you don’t want a local dev environment, or for quick one-off edits and chat-augmented exploration.

All five approaches can be combined. A common pattern is developing workflows locally with the SDK, mutating entities (forms, agents, configs) via CLI, and using MCP for exploration.

AI-Assisted Development

1. Claude Code with the Bifrost plugin

Install the plugin

2. Codex with the Bifrost plugin

3. Other Coding Agents with bifrost skill

Migrating from a hand-copied skill

Setup after installing a plugin or skill

How it works

CLI commands for platform operations

When to use which

4. Local SDK Development

Setup

Workflow

What to tell your AI tool

5. MCP for External AI Tools

Setup

What MCP gives you

MCP system prompt

Choosing an Approach

3. Other Coding Agents with `bifrost skill`