streaming

streaming — built-in tools + terminal chat rendering

What this shows

An agent wired with the standard built-in tools (read_file, list_dir, bash, ...) behind a permissions gate, driven from a command-line prompt and rendered as an interactive chat session via runner.WriteEvents. When stdout is a terminal, tool calls render in cyan and partial assistant text in green; pipe the output and the same code path emits plain text with no escape codes.

The permissions mode is set to yolo so tool calls auto-approve — the example has no human in the loop. Real apps keep ask and wire a prompter into permissions.FromConfig.

Run it

Needs real provider credentials; the provider auto-detects from env when --provider is omitted:

GEMINI_API_KEY=...    go run ./examples/streaming "what's in main.go in this directory?"
ANTHROPIC_API_KEY=... go run ./examples/streaming --provider anthropic "list the .go files here"

Pick a prompt that forces tool use ("how many lines of go code in this repo?" → bash) to see the formatter work.

Key APIs

• runner.WriteEvents / runner.WithColor / runner.IsTerminal — github.com/go-steer/core-agent/v2/pkg/runner
• tools.Build + tools.Default — github.com/go-steer/core-agent/v2/pkg/tools
• permissions.FromConfig / permissions.ModeYolo — github.com/go-steer/core-agent/v2/pkg/permissions
• agent.New / agent.WithTools — github.com/go-steer/core-agent/v2/pkg/agent

What you should see

> what's in main.go in this directory?
→ read_file {"path": "main.go"}
main.go is the streaming example: it builds the default toolset, ...

with the → tool lines colored when run in a terminal.

Next

• with-tools — define your own tool and load MCP servers + skills.
• replay — run the same event loop offline against a recorded transcript.
• Library guide: docs/site/src/content/docs/embed/guide.md