bridle

An agentic harness and SDK for OpenAI-compatible chat completion endpoints.

View it on OpenRouter

Features

Harness user experience:

• Readline-based CLI with tab completion
  • Coloured output for sections (-Colour, or NO_COLOR env var)
  • Description of every available tool via /tools
  • Press Ctrl+X to edit the current input line in $EDITOR
  • Press any key to interrupt a running model turn or tool call
  • Audible bell after 10 seconds to signal the terminal is ready for input
• Oneshot mode (-p, implies -Agent=oneshot)
• Select the agent logic at startup (-Agent=oneshot|interactive|autoengineer)
• Manipulate turn history (/rewind, /turns, /submit)
  • /rewind prompts for a target index from /turns, defaults to today
• Session management (-Autosave, manual /save, and resume via -r)
• Import local files (wildcard match) into prompt (/read)
• Restart the session with a freshly compiled binary (/reexec)
• Linux sandboxing with Bubblewrap, on by default (-Sandbox=false to opt out, use at own risk)
  • Append extra Bubblewrap flags to every sandboxed command via -SandboxFlag (can be specified multiple times)
  • Leaves the harness process unconstrained and confines each thing the model can reach: file access by path, shell commands and MCP stdio servers by their own namespaces
  • Grant extra paths with -Allow, applied to every confined operation (repeatable)
    • A directory exposes itself and everything inside it; a single file exposes just that file
• Git worktree support (-w), also works from inside the sandbox
• Quiet mode (-Quiet) to suppress all system messages

Special modes:

• "Autoengineer" (-Agent=autoengineer)
  • Implement every task in TODO file and TODO/FIXME comments in codebase.
  • The result is not necessarily very good, but it can run overnight and you can cherry-pick the good parts in the morning.

Model integration with harness:

• Auto load AGENTS.md and CLAUDE.md files
• Agent Skills support
  • Discovers SKILL.md bundles under .agents/skills/ or .claude/skills/
  • Scanned at project level and user level, project wins on name collision
  • The skill catalog is injected into the system prompt
  • Use read_skill to load the full contents of a skill (no size limit)
• Context management
  • Auto compact on client-side token threshold (-MaxContext) or on server-side length exceeded
  • Soft limit: compaction fires 10% below the ceiling to leave a safety buffer
  • Auto-detect model context from /v1/models when -MaxContext is zero
  • Several strategies behind one dispatcher: handover (default), head-summarize, quick, summarize, drop-reasoning, drop-tool-calls, truncate, rewind
  • The default "handover" strategy appends a summary instead of rewriting history, preserving prefix caching
• Tool calling
  • Available tools:
    • Local file I/O (read_file, write_file), with optional limit/offset to read a line range of a large file
    • Single-string replacement (edit_file) for reliable targeted edits
    • Auto-healing apply_patch, supporting both unified + V4A formats
    • Command execution (local_shell)
    • Subagents (subagent with fork/fresh context modes), optionally running in parallel (-Parallel <n>)
    • Web search with DuckDuckGo
    • Webpage read in Reader Mode (via go-readability), converting links to navigable markdown
    • Anti-hallucination: write_file/apply_patch must read a file first (or be nonexistent) before writing, and web_read only permits URLs already seen in the conversation
  • MCP support
    • Stdio, Streaming HTTP, or SSE transports
    • Static (-Mcp <name>=<command> / -Mcp <name>=http://url) or dynamic registration (/mcp-connect, /mcp-list, /mcp-disconnect)
    • Dynamic tool discovery
    • When using sandboxing, stdio servers are confined like any other subprocess (with the host network, so they can still reach an API); HTTP servers are remote and have no local process to confine
  • User confirmation levels (/checktools or -CheckTools: if-sandboxed default, ask, auto, or yolo)
    • if-sandboxed becomes yolo when the tools are confined and ask when bubblewrap is unavailable
    • Allow denying with alternative message (?)
    • Refuse and stop the turn outright (!), as distinct from n, which refuses this call but lets the model try something else
  • Trigger tools locally (/tool), optionally submit in next turn

Model settings:

• Use custom model provider (-Host and -Key)
  • Supports timing metadata from both llama-server and OpenAI/OpenRouter
  • Can be changed mid-session with /host, /key and /model
• Merge a custom JSON object into each API request body (-ExtraArgs), e.g. OpenRouter provider preferences
• Toggle thinking (/think and /no-think) and set reasoning_effort (/reasoning-effort)
• Running cost is tracked and reported per turn when the provider reports it
• Customizable system message, default-tuned for Golang development

Usage

# For 128k 131072, for 256k 262144 . Mostly affects pp speed, not tg speed
llama-server -c 262144 --cache-type-k q8_0 --cache-type-v q8_0 -fa on -np 1 -m /srv/llama/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf &

./bridle

# or
./bridle -CheckTools=yolo -Autosave=auto -w my-worktree

# Interactive session, with up to 4 parallel subagents
./bridle -Parallel 4

# Use with OpenRouter
./bridle -Host 'https://openrouter.ai/api/' -Key 'sk-or-v1-xxxx' -Model 'openrouter/free'