---

grip is a self-hostable AI agent platform — 120+ Python modules, ~24,000 lines, 826 tests. It uses the Claude Agent SDK as its primary engine for Claude models, with a LiteLLM fallback for 15+ other providers (OpenAI, DeepSeek, Groq, Gemini, Ollama local & cloud, etc.). Chat over Telegram/Discord/Slack, automate with a headless browser, track multi-step tasks, schedule reliable cron jobs, orchestrate multi-agent workflows, and expose a secure REST API — all from a single grip gateway process.

Features

Category	Details
Dual Engine	Claude Agent SDK (primary, recommended) + LiteLLM fallback for non-Claude models
LLM Providers	Anthropic (via SDK), OpenRouter, OpenAI, DeepSeek, Groq, Google Gemini, Qwen, MiniMax, Moonshot (Kimi), Ollama (Cloud), Ollama (Local), vLLM, Llama.cpp, LM Studio, and any OpenAI-compatible API
Built-in Tools	31 tools across 16 modules — file read/write/edit/append/list/delete, shell execution, web search (Brave + DuckDuckGo), deep web research, browser automation (Playwright — navigate, click, fill, screenshot, evaluate JS, extract content), document conversion (MarkItDown — PDF/DOCX/PPTX/XLSX/HTML/images), code analysis, data transforms, document generation, email composition, task tracking (todo_write/todo_read), messaging, subagent spawning, finance (stock quotes, history, company info via yfinance), cron scheduling, workflows, MCP tools
Task Tracking	todo_write/todo_read tools with workspace persistence — active tasks injected into every system prompt so the agent never loses track across iterations
Chat Channels	Telegram (bot commands, photos, documents with auto-conversion, voice), Discord (attachments with auto-conversion), Slack (Socket Mode, file shares with auto-conversion)
REST API	FastAPI with bearer auth, rate limiting, audit logging, security headers, 28 endpoints
Workflows	DAG-based multi-agent orchestration with dependency resolution and parallel execution
Memory	Dual-layer (MEMORY.md + HISTORY.md) with hybrid search (FTS5 BM25 + vector embeddings merged via Reciprocal Rank Fusion), TF-IDF fallback, auto-consolidation, mid-run compaction, semantic caching, and knowledge base
Scheduling	Cron jobs with state machine (pending→fired→running→succeeded/failed), idempotency keys, deferred retry queue, channel delivery, heartbeat service, natural language scheduling
Skills	15 built-in markdown skills, workspace overrides, install/remove via CLI
Security	Directory trust model, shell deny-list (50+ patterns), credential scrubbing, Unicode sanitization (lone surrogate stripping), SecretStr config fields, secret sanitizer, Shield runtime threat feed policy, token tracking, rate limiting
OAuth	Auto-refresh with 5-minute proactive window, concurrent refresh protection, per-server token isolation, PKCE authorization code flow
Observability	OpenTelemetry tracing, in-memory metrics, crash recovery, config validation

Architecture

grip gateway
├── REST API (FastAPI :18800)          28 endpoints, bearer auth, rate limiting
│   ├── /api/v1/chat                   blocking + SSE streaming
│   ├── /api/v1/sessions               CRUD
│   ├── /api/v1/tools                  list + execute
│   ├── /api/v1/convert                document-to-markdown file upload
│   ├── /api/v1/mcp                    server management + OAuth
│   └── /api/v1/management             config, cron, skills, memory, metrics, workflows
├── Channels
│   ├── Telegram                       bot commands, photos, docs (auto-convert), voice
│   ├── Discord                        discord.py + attachment auto-convert
│   └── Slack                          Socket Mode + file share auto-convert
├── Message Bus                        asyncio.Queue decoupling channels ↔ engine
├── Engine (pluggable)
│   ├── SDKRunner (claude_sdk)         Claude Agent SDK — full agentic loop
│   └── LiteLLMRunner (litellm)        any model via LiteLLM + grip's AgentLoop
├── Tool Registry                      31 tools across 18 modules
│   ├── filesystem                     read/write/edit/append/list/delete/trash
│   ├── shell                          exec with 50+ pattern deny-list
│   ├── web                            web_search + web_fetch (HTML→markdown)
│   ├── browser                        Playwright headless Chromium (navigate/click/fill/screenshot/evaluate/content)
│   ├── research                       deep web_research (HTML→markdown)
│   ├── markitdown                     convert_document (PDF/DOCX/XLSX/images/…)
│   ├── message                        send_message + send_file
│   ├── spawn                          subagent spawn/check/list
│   ├── todo                           todo_write + todo_read (task tracking)
│   ├── workflow                       multi-agent DAG execution
│   ├── scheduler                      cron scheduling
│   ├── finance                        stock_quote, stock_history, company_info (yfinance)
│   └── mcp                            MCP tool proxy
├── MCP Manager                        stdio + HTTP/SSE servers, OAuth 2.0 + PKCE
├── Memory
│   ├── MEMORY.md                      durable facts (TF-IDF search, Jaccard dedup)
│   ├── HISTORY.md                     timestamped summaries (time-decay search)
│   ├── HybridSearch                   FTS5 BM25 + vector cosine, merged via RRF
│   ├── brain.db                       SQLite FTS5 index + vector embeddings
│   ├── SemanticCache                  SHA-256 keyed response cache with TTL
│   └── KnowledgeBase                  structured typed facts
├── Session Manager                    per-key JSON files, LRU cache (200)
├── Cron Service                       state machine, idempotency, deferred retry, channel delivery
├── Heartbeat Service                  periodic autonomous agent wake-up
└── Workflow Engine                    DAG execution with topological parallelism

Engine Modes

grip uses a dual-engine architecture controlled by the engine config field:

Engine	Config Value	Use Case
Claude Agent SDK	claude_sdk (default)	Anthropic Claude models — full agentic loop, tool execution, and context management handled by the SDK
LiteLLM	litellm	Non-Claude models (OpenAI, DeepSeek, Groq, Gemini, Ollama Cloud/Local, etc.) — uses grip's internal agent loop with LiteLLM for API calls

Switch engines via config:

# Use Claude Agent SDK (default)
grip config set agents.defaults.engine "claude_sdk"

# Use LiteLLM for non-Claude models
grip config set agents.defaults.engine "litellm"

The onboarding wizard (grip onboard) automatically sets the right engine based on your provider choice.

Installation

Recommended: Via PyPI

# Using uv (faster)
uv tool install grip-ai

# Using pip
pip install grip-ai

Manual Install (from source)

# Clone the repository
git clone https://github.com/5unnykum4r/grip-ai.git
cd grip-ai

# Install (includes Telegram, REST API, LiteLLM, and all core features)
uv sync

# Optional extras
uv sync --extra discord      # Discord bot
uv sync --extra slack        # Slack bot (Socket Mode)
uv sync --extra mcp          # Model Context Protocol
uv sync --extra document     # Document conversion (MarkItDown — PDF/DOCX/XLSX/images)
uv sync --extra browser      # Browser automation (Playwright — headless Chromium)
uv sync --extra viz          # Data visualization (plotext)
uv sync --extra observe      # OpenTelemetry tracing
uv sync --extra all          # Everything above

# Register grip command globally (development/editable mode)
uv tool install --editable .

Requirements

• Python 3.12+
• uv package manager
• An Anthropic API key (for Claude Agent SDK) or an API key from another LLM provider (with litellm engine)

Quickstart

1. Run the Setup Wizard

grip onboard

The wizard walks you through:

• Choosing your engine: Claude Agent SDK (recommended) or LiteLLM (15+ providers)
• Entering your API key
• Choosing a default model
• Configuring hybrid search (embedding model for semantic memory retrieval)
• Optional Telegram bot setup
• File access mode (workspace-only, prompt, or trust-all)
• Testing connectivity

2. Chat with the Agent

# Interactive mode
grip agent

# One-shot message
grip agent -m "What files are in my workspace?"

# Pipe input from stdin
cat error.log | grip agent -m "Fix this error"

Interactive mode supports slash commands:

Command	Description
/new	Start a fresh conversation
/clear	Clear conversation history
/undo	Remove last exchange
/rewind N	Rewind N exchanges
/compact	Compress session history
/copy	Copy last response to clipboard
/model <name>	Switch AI model
/provider	Show current provider details
/tasks	Show scheduled cron tasks
/trust <path>	Grant agent access to a directory
/trust revoke <path>	Revoke agent access to a directory
/status	Show session info
/mcp	List MCP servers
/doctor	Run diagnostics
/help	List all commands
/exit	Exit grip

3. Start the Gateway

The gateway is the long-running process that connects everything — Telegram, cron, heartbeat, and the REST API:

grip gateway

Telegram Setup

Step 1: Create a Bot

1. Open Telegram and search for @BotFather
2. Send /newbot and follow the prompts
3. Copy the bot token (looks like 123456:ABC-DEF1234ghIkl-zyx57W2v1u123ew11)

Step 2: Configure grip

# Set your bot token
grip config set channels.telegram.enabled true
grip config set channels.telegram.token "YOUR_BOT_TOKEN"

Optional: Restrict the bot to specific Telegram user IDs for security:

# Find your Telegram user ID by messaging @userinfobot
grip config set channels.telegram.allow_from '["YOUR_TELEGRAM_USER_ID"]'

Step 3: Start the Gateway

grip gateway

You should see:

Channels started: telegram
API server started: http://127.0.0.1:18800
grip gateway running. Press Ctrl+C to stop.

Step 4: Chat with Your Bot

Open Telegram and message your bot. Available bot commands:

Command	Description
/start	Welcome message
/help	List commands
/new	Start fresh conversation
/status	Session info (model, message count, memory)
/model <name>	Switch AI model (e.g. /model openai/gpt-4o)
/trust <path>	Grant agent access to a directory (e.g. /trust ~/Downloads)
/trust revoke <path>	Remove a directory from the trusted list
/undo	Remove last exchange
/clear	Clear conversation history
/compact	Summarize and compress session

Send any text message to chat with the AI. The bot also handles photos (with OCR/description via MarkItDown), documents (auto-converted to markdown — PDF, DOCX, XLSX, etc.), and voice messages.

Working with Telegram

Setting Reminders:

Tell the agent to remind you of something — it creates a cron job and delivers the result back to your Telegram chat:

"Remind me to check the server status in 30 minutes"

The agent will create a cron job with --reply-to pointed at your Telegram chat, so the reminder is delivered directly to you.

Switching Models:

/model openrouter/google/gemini-2.5-pro

Starting Fresh:

/new

Compressing Long Sessions:

When a conversation gets long, compress it to save tokens:

/compact

CLI Commands

Command	Description
grip onboard	Interactive setup wizard
grip agent	Chat with the AI agent (interactive or one-shot)
grip gateway	Run full platform: API + channels + cron + heartbeat
grip serve	Start standalone REST API server
grip status	Show system status
grip update	Pull latest source and re-sync dependencies
grip config	View and modify configuration (show, set, path)
grip cron	Manage scheduled jobs (list, add, remove, enable, disable)
grip skills	Manage agent skills (list, install, remove)
grip workflow	Manage multi-agent workflows (list, show, run, create, delete)
grip mcp	Manage MCP server configurations (list, add, remove, presets)

Global flags: --verbose / -v, --quiet / -q, --config / -c PATH, --dry-run.

Configuration

Config is stored at ~/.grip/config.json. Environment variables override with GRIP_ prefix and __ nested delimiter.

# View current config (secrets are masked)
grip config show

# Set values
grip config set agents.defaults.model "anthropic/claude-sonnet-4"
grip config set agents.defaults.max_tokens 16384
grip config set agents.defaults.temperature 0.7

# Unlimited tool iterations (default) — agent stops naturally when done
grip config set agents.defaults.max_tool_iterations 0

# Cap tool iterations (e.g. safety limit for automated tasks)
grip config set agents.defaults.max_tool_iterations 50

Key Sections

Section	Description
agents.defaults	Engine (claude_sdk/litellm), SDK model, default model, max_tokens, temperature, memory_window, max_tool_iterations (0=unlimited), workspace path
agents.defaults.search	Hybrid search: embedding_model, embedding_dimensions, vector_weight, bm25_weight, rrf_k, auto_reindex
agents.profiles	Named agent configs (model, tools_allowed, tools_denied, system_prompt_file)
agents.model_tiers	Cost-aware routing: different models for low/medium/high complexity prompts
providers	Per-provider API keys and base URLs
tools	Web search config, shell_timeout, workspace sandboxing, MCP servers
channels	Telegram/Discord/Slack tokens, allow_from lists
gateway	Host, port, API auth, rate limits, CORS, request size limits
heartbeat	Periodic autonomous agent runs (enabled, interval_minutes)
cron	Scheduled task settings (exec_timeout_minutes)

Agent Profiles

Define specialized agents for different tasks:

{
  "agents": {
    "profiles": {
      "researcher": {
        "model": "openai/gpt-4o",
        "tools_allowed": ["web_search", "web_fetch"],
        "system_prompt_file": "RESEARCHER.md"
      },
      "coder": {
        "model": "anthropic/claude-sonnet-4",
        "tools_denied": ["exec"],
        "temperature": 0.3
      }
    }
  }
}

Cost-Aware Model Routing

Route prompts to different models based on complexity:

grip config set agents.model_tiers.enabled true
grip config set agents.model_tiers.low "openrouter/google/gemini-flash-2.0"
grip config set agents.model_tiers.high "openrouter/anthropic/claude-sonnet-4"

Simple queries (greetings, lookups) go to the cheap model. Complex tasks (architecture, debugging) go to the powerful model.

Task Tracking

For multi-step tasks, the agent maintains a persistent task list in workspace/tasks.json. Active tasks are automatically injected into every system prompt so the agent always knows where it left off — even across long runs with context compaction.

The agent uses two built-in tools:

Tool	Description
todo_write	Create or replace the full task list (persisted to workspace/tasks.json)
todo_read	Read the current task list with statuses

Example — how the agent handles a big task:

User: "Build me a REST API with auth, CRUD for users, and tests"

Agent:
  1. Calls todo_write to create the task plan:
     ○ [1] Design data models and schema — pending
     ○ [2] Implement auth endpoints — pending
     ○ [3] Implement user CRUD endpoints — pending
     ○ [4] Write tests — pending

  2. Updates status before each step:
     ◑ [1] Design data models and schema — in_progress

  3. Marks done, moves to next:
     ● [1] Design data models and schema — completed
     ◑ [2] Implement auth endpoints — in_progress

The task list is visible in ~/.grip/workspace/tasks.json and cleared/updated as the agent progresses.

MCP Servers

grip supports Model Context Protocol servers via stdio or HTTP/SSE transport.

# List available presets
grip mcp presets

# Add preset servers
grip mcp presets todoist excalidraw firecrawl

# Add all presets
grip mcp presets --all

# Add custom HTTP server
grip mcp add myserver --url https://mcp.example.com

# Add custom stdio server
grip mcp add myserver --command npx --args -y,mcp-server

# List configured servers
grip mcp list

# Remove a server
grip mcp remove excalidraw

Available Presets (14)

Name	Type	Description
todoist	stdio	Task management
excalidraw	HTTP	Collaborative whiteboard
firecrawl	stdio	Web scraping (requires API key)
bluesky	stdio	Social network
filesystem	stdio	File system access
git	stdio	Git operations
memory	stdio	Knowledge persistence
postgres	stdio	PostgreSQL queries
sqlite	stdio	SQLite database
fetch	stdio	HTTP fetching
puppeteer	stdio	Browser automation
stack	stdio	Stack Overflow Q&A
tomba	stdio	Email finder (Tomba.io)
supabase	HTTP	Supabase database + auth

Set API keys for presets that require them:

grip config set tools.mcp_servers.firecrawl.env.FIRECRAWL_API_KEY "your-key"

Workflows

Create multi-agent workflows as JSON files:

{
  "name": "research-and-summarize",
  "description": "Research a topic and produce a summary",
  "steps": [
    {
      "name": "research",
      "prompt": "Research the latest developments in quantum computing",
      "profile": "researcher",
      "timeout_seconds": 600
    },
    {
      "name": "summarize",
      "prompt": "Summarize the following research: {{research.output}}",
      "profile": "coder",
      "depends_on": ["research"]
    }
  ]
}

grip workflow create research.json
grip workflow run research-and-summarize
grip workflow list

Steps with no dependencies execute in parallel. The engine uses Kahn's algorithm for topological ordering and detects cycles at validation time.

Cron & Scheduling

Schedule tasks and reminders with cron expressions or natural language:

# Add a reminder
grip cron add "standup" "0 9 * * 1-5" "Remind the user: standup in 15 minutes"

# Add a task that reports to Telegram
grip cron add "disk-check" "0 */6 * * *" "Check disk usage" --reply-to "telegram:YOUR_CHAT_ID"

# Manage jobs
grip cron list
grip cron disable <job-id>
grip cron enable <job-id>
grip cron remove <job-id>

When chatting via Telegram/Discord/Slack, the agent automatically sets --reply-to so cron results are delivered to your chat.

Scheduler Reliability

The cron service uses a state machine to track job lifecycle and prevent common scheduling pitfalls:

• State machine: Each job transitions through pending → fired → running → succeeded/failed — last_run is set AFTER execution, not before, so a job can't appear "done" when it never ran
• Idempotency keys: Jobs are deduplicated by content hash (name + schedule + prompt). Creating the same logical job twice returns the existing one instead of accumulating duplicates
• Deferred retry queue: If a job fires while the engine is busy, it enters a retry queue instead of being silently dropped. Retried on the next cycle with a 10-minute max age
• Graceful shutdown: The service waits for in-flight jobs to complete before stopping

API Reference

Start the API server standalone or as part of the gateway:

# Standalone
grip serve

# Full gateway (API + channels + cron + heartbeat)
grip gateway

On first run, an auth token is auto-generated and printed to stderr.

Endpoints

All /api/v1/* endpoints require Authorization: Bearer <token>.

Method	Path	Description
GET	/health	Load balancer probe (no auth)
GET	/api/v1/health	Authenticated health with version + uptime
POST	/api/v1/chat	Send message, get response
GET	/api/v1/sessions	List session keys
GET	/api/v1/sessions/{key}	Session detail
DELETE	/api/v1/sessions/{key}	Delete session
GET	/api/v1/tools	List tool definitions
POST	/api/v1/tools/{name}/execute	Execute tool directly (disabled by default)
POST	/api/v1/convert	Convert uploaded document to markdown (MarkItDown)
GET	/api/v1/status	System status
GET	/api/v1/config	Masked config dump
GET	/api/v1/metrics	Runtime metrics
GET	/api/v1/cron	List cron jobs
POST	/api/v1/cron	Create cron job
DELETE	/api/v1/cron/{id}	Delete cron job
POST	/api/v1/cron/{id}/enable	Enable cron job
POST	/api/v1/cron/{id}/disable	Disable cron job
GET	/api/v1/skills	List loaded skills
GET	/api/v1/memory	Read MEMORY.md
GET	/api/v1/memory/search?q=...	Search HISTORY.md
GET	/api/v1/workflows	List workflows
GET	/api/v1/workflows/{name}	Workflow detail
GET	/api/v1/mcp/servers	List MCP servers with status
GET	/api/v1/mcp/{server}/status	Single server status
POST	/api/v1/mcp/{server}/login	Start OAuth flow
GET	/api/v1/mcp/callback	OAuth redirect handler (no auth)
POST	/api/v1/mcp/{server}/enable	Enable a server
POST	/api/v1/mcp/{server}/disable	Disable a server

Example Requests

# Health check
curl http://localhost:18800/health

# Chat
curl -X POST http://localhost:18800/api/v1/chat \
  -H "Authorization: Bearer grip_YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"message": "Hello, what can you do?"}'

# List tools
curl -H "Authorization: Bearer grip_YOUR_TOKEN" http://localhost:18800/api/v1/tools

# Metrics
curl -H "Authorization: Bearer grip_YOUR_TOKEN" http://localhost:18800/api/v1/metrics

# Convert a document to markdown
curl -X POST http://localhost:18800/api/v1/convert \
  -H "Authorization: Bearer grip_YOUR_TOKEN" \
  -F "file=@document.pdf"

Security

The API is designed for safe self-hosting:

• Bearer token auth with timing-safe comparison
• Rate limiting — per-IP (30/min) + per-token (60/min), sliding window
• Request size limit — 1 MB default, rejected before parsing
• Security headers — X-Content-Type-Options, X-Frame-Options, Content-Security-Policy
• Audit logging — every request logged (method, path, status, duration, IP)
• Directory Trust Model — grip is restricted to its workspace by default. Access to external directories must be explicitly granted via CLI prompt or the /trust command. Learn more.
• Shell Safety Guards — every shell command is scanned against a comprehensive deny-list (50+ patterns) before execution. Learn more.
• Credential Scrubbing — API keys, tokens, and passwords in tool outputs are automatically redacted before being stored in message history.
• Shield Policy — context-based runtime threat feed injected into the system prompt. Evaluates tool calls, skill execution, MCP interactions, network requests, and secret access against active threats. Learn more.
• SecretStr config fields — API keys and tokens use Pydantic SecretStr, automatically masked in logs and repr() output
• Sanitized errors — no stack traces or file paths in responses
• Tool execute gated — disabled by default to prevent arbitrary command execution over HTTP
• No config mutation over HTTP — prevents redirect attacks
• Startup warnings — alerts for dangerous configs (0.0.0.0 binding, tool execute enabled)

Security Architecture

Grip implements a multi-layered defense to make the platform as safe as possible for self-hosting.

1. Directory Trust Model

Grip is sandboxed to its workspace by default. Unlike traditional agents with unrestricted disk access, Grip cannot touch your personal files unless you "opt-in".

• Workspace First: The agent can always read/write within its assigned workspace directory.
• Explicit Consent: To access any directory outside the workspace (e.g., ~/Downloads), the user must explicitly "trust" it.
• Persistent Safety: Trust decisions are saved and remembered across sessions.

2. Shield Policy (Runtime Threat Feed)

The agent's system prompt includes a SHIELD.md policy that defines how to evaluate actions against a threat feed before execution. This is a context-level safety layer that works alongside the code-level guards.

Scopes covered: prompt, skill.install, skill.execute, tool.call, network.egress, secrets.read, mcp

Enforcement actions:

• block — stop immediately, no execution
• require_approval — ask the user for confirmation before proceeding
• log — continue normally (default when no threat matches)

How it works:

1. Active threats are injected into the ## Active Threats section of SHIELD.md at runtime
2. Before acting on a scoped event, the agent evaluates it against loaded threats
3. Matching uses category/scope alignment, recommendation_agent directives, and exact string fallback
4. Confidence threshold (>= 0.85) determines enforceability; below that, the action defaults to require_approval
5. When multiple threats match, the strictest action wins: block > require_approval > log

The policy is stored at workspace/SHIELD.md and can be customized per workspace.

3. Shell Command Deny-List

Every command the agent tries to run via the exec tool is scanned against a robust list of 50+ dangerous patterns before execution:

• Destructive Commands: Blocked rm -rf /, rm -rf ~, mkfs, etc.
• System Control: Blocked shutdown, reboot, systemctl poweroff.
• Credential Exfiltration: Blocked cat ~/.ssh/id_rsa, cat .env, etc.
• Remote Code Injection: Blocked curl | bash and similar pipe-to-shell patterns.

4. Credential Scrubbing

Tool outputs are automatically scrubbed before being stored in message history. Patterns detected and redacted:

• sk-... API keys (OpenAI/Anthropic-style)
• ghp_... GitHub personal access tokens
• xoxb-... Slack bot tokens
• Bearer <token> authorization headers
• password=... URL and config parameters

[!IMPORTANT] While we strive for "perfect safety" through these multi-layered guards, no system is infallible. Always run grip with a non-root user and review critical actions.

Why it's better:

• Zero-Trust for sensitive data: Even top-tier LLMs can "hallucinate". Our guards make it physically impossible for the agent to exfiltrate your SSH keys or delete your home folder by accident.
• Controlled Blast Radius: By restricting the agent to specific folders, you ensure an accidental "delete all" command only affects the project directory you're working in.
• Privacy by Design: You maintain absolute control over the agent's data footprint.

Managing Trust:

• In Chat (Telegram/Discord/Slack):
  • /trust <path> — Grant permanent access to a directory.
  • /trust revoke <path> — Remove access for a directory.
  • /trust — List all currently trusted directories.
• In CLI interactive mode: Same /trust commands work in grip agent.
• Manual Control: Trust decisions are stored in your_workspace/state/trusted_dirs.json. You can manually edit or clear this file to manage access at any time.

Long-Running Tasks

grip is designed to handle complex, multi-step work without hitting context limits.

Unlimited Iterations

By default, max_tool_iterations = 0 (unlimited). The agent runs until it has nothing left to do — no artificial cap. For automated jobs where you want a safety limit:

grip config set agents.defaults.max_tool_iterations 100

Mid-Run Compaction

When in-flight messages exceed 50, the agent automatically LLM-summarizes the older ones and compacts them into a single summary block, keeping the 20 most recent messages intact. This prevents context overflow on long tasks (building full websites, large refactors, deep research) without losing continuity.

The compaction is triggered mid-iteration — transparent to the user. A consolidation model can be configured to save tokens:

grip config set agents.defaults.consolidation_model "openrouter/google/gemini-flash-2.0"

Task Persistence

The agent creates a tasks.json in the workspace at the start of any multi-step task. If a session is interrupted or compacted, the task list is re-injected into the system prompt at the next iteration, so the agent picks up exactly where it left off.

Browser Automation

grip includes a built-in Playwright-powered browser tool that gives the agent full headless Chromium control — handling JavaScript-rendered SPAs, login flows, and dynamic content that web_fetch can't reach.

Setup

# Install the browser extra
uv sync --extra browser

# Install Chromium (one-time)
playwright install chromium

Actions

Action	Description
navigate	Go to a URL, returns page title and HTTP status
click	Click an element by CSS selector
fill	Type text into an input field
select	Pick an option from <select> dropdowns by value or visible label
press	Press a keyboard key (Enter, Escape, Tab, etc.), optionally on a focused element
scroll	Scroll the page up or down by pixel amount (default 500px)
screenshot	Capture the page — save to file or return base64
content	Extract rendered text from <main>, <article>, or <body>
evaluate	Run arbitrary JavaScript and return the result
wait	Wait for a CSS selector to become visible
back	Navigate back in browser history
close	Close the browser session

Chromium is auto-installed on first use if the binary is missing. The browser session is reused across calls — the agent can navigate, interact, and extract data across multiple tool invocations without re-launching Chromium each time.

Docker

Grip is Docker-ready and can be configured entirely via environment variables.

# Build from source
docker build -t grip .

# Run with Claude Agent SDK (recommended)
docker run -d \
  -p 18800:18800 \
  -e ANTHROPIC_API_KEY="sk-ant-..." \
  -e GRIP_CHANNELS__TELEGRAM__ENABLED="true" \
  -e GRIP_CHANNELS__TELEGRAM__TOKEN="bot-token" \
  -v ~/.grip:/home/grip/.grip \
  --name grip-agent \
  grip

# Run with LiteLLM engine (non-Claude models)
docker run -d \
  -p 18800:18800 \
  -e GRIP_AGENTS__DEFAULTS__ENGINE="litellm" \
  -e GRIP_PROVIDERS__OPENAI__API_KEY="sk-..." \
  -e GRIP_CHANNELS__TELEGRAM__ENABLED="true" \
  -e GRIP_CHANNELS__TELEGRAM__TOKEN="bot-token" \
  -v ~/.grip:/home/grip/.grip \
  --name grip-agent \
  grip

Configuration via Environment

Grip supports GRIP_ prefixed variables for any config value. Use __ for nested keys:

• GRIP_AGENTS__DEFAULTS__MODEL
• GRIP_PROVIDERS__ANTHROPIC__API_KEY
• GRIP_GATEWAY__PORT

Built-in Skills

Skill	Description	Always Loaded
code-review	Automated code review and quality analysis	Yes
optimization-rules	Token efficiency and tool selection guidance	Yes
code-loader	AST-aware chunking for loading relevant code
codebase-mapper	Dependency graphs, import mapping, ripple analysis
data-viz	ASCII charts and data visualization in terminal
debug	Bug finding, git blame, bisect, time-travel debugging
github	PR generation, code review, git workflows
memory	Long-term memory management
project-planner	Project planning and task breakdown
self-analyzer	Performance and architecture analysis
skill-creator	Create new skills
summarize	Text and conversation summarization
temporal-memory	Time-aware reminders and deadline tracking
tmux	Terminal multiplexer management
tweet-writer	Social media content drafting

Development

# Install dev dependencies
uv sync --group dev

# Run linter
uv run ruff check grip/ tests/

# Run tests (826 tests across 50+ test files)
uv run pytest

# Run tests with coverage
uv run pytest --cov=grip

# Run specific test module
uv run pytest tests/memory/ -v

# Build package
uv build

Project Stats

Metric	Count
Python source files	120+
Lines of code	~24,000
Tests	826
Built-in tools	31 (16 modules)
Built-in skills	15
LLM providers	15
API endpoints	28
CLI commands	11 groups + 16 interactive slash commands

Contributing

Contributions are welcome! Here's how to get started:

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Make your changes
4. Run linting and tests: uv run ruff check grip/ tests/ && uv run pytest
5. Commit your changes: git commit -m "Add my feature"
6. Push to your fork: git push origin feature/my-feature
7. Open a Pull Request

Please ensure:

• All tests pass (uv run pytest)
• No lint errors (uv run ruff check grip/ tests/)
• New features include tests where appropriate

License

MIT

Disclaimer

grip is an AI-powered platform designed for high autonomy. Please be aware:

• AI models can produce hallucinations, errors, or unexpected outputs.
• Autonomous tool execution (especially shell/exec) carries inherent security risks.
• Users are responsible for monitoring agent behavior and ensuring compliance with LLM provider terms.
• Use this software at your own risk.