Tapir

About

Tapir is an AI agent coding assistant that runs in your terminal. Built in Rust, it connects to the Anthropic API to orchestrate AI agents through complex software engineering tasks.

Tapir is probably not for you if you think a good tool needs an Electron app, a plugin marketplace, and a few gigabytes of RAM. The binary is under 1MB. At runtime it uses a few megabytes of resident memory — not gigabytes:

Tapir does not ask for permission before executing tools — it reads, writes, and runs commands as instructed by the model. The recommended way to run it is inside an isolated environment such as a container, so the agent can operate freely without risk to your host system.

All four tools below were given the same task: "read @Cargo.toml and return the dependencies list". Here is how much memory each one used:

No frameworks. No garbage collector. No runtime. Just ~5500 lines of Rust that talk to an API and edit your files.

Features

• Streaming responses — real-time token output with a thinking timer
• Tool execution — read, write, edit files, bash, ls, find, grep — run in parallel
• Session management — persist and resume conversations across sessions
• Context files — auto-load AGENTS.md, CLAUDE.md, and custom system prompts
• Skill system — discover and load skills from project, user, and global directories
• Conversation compaction — automatic summarization when context gets large
• Extended thinking — configurable thinking budget for complex reasoning
• Shell escapes — !cmd sends output to the LLM, !!cmd runs locally
• Ctrl+C handling — interrupt responses or tool execution cleanly
• Collapsible tool output — toggle with Ctrl+O to keep your terminal clean
• Tab completion — complete @file paths in the input line
• HTTP retry with backoff — automatic retry on transient API errors
• Static binary — single musl binary under 900KB, zero dependencies

Setup

Tapir connects to the Anthropic API by default. Set your API key as an environment variable:

export ANTHROPIC_API_KEY="sk-ant-..."

To use a different provider or endpoint (e.g. a local proxy), override the API URL:

export TAPIR_API_URL="https://your-provider.example.com/v1/messages"

You can also configure the model and token limits:

ANTHROPIC_API_KEY

API key (required)

TAPIR_MODEL

Model ID (default: claude-opus-4-6)

TAPIR_MAX_TOKENS

Max response tokens (default: 16384)

TAPIR_THINKING_BUDGET

Extended thinking budget, 0 = off (default: 0)

TAPIR_API_URL

API endpoint (default: https://api.anthropic.com/v1/messages)

Environment variables take priority over the config file.

Configuration

Tapir reads a JSON config file from ~/.tapir/config.json. You can specify a custom path with tapir -c /path/to/config.json.

{
  "api_key": "sk-ant-...",
  "model": "claude-sonnet-4-6",
  "max_tokens": 8192,
  "thinking_budget": 10000,
  "api_url": "https://api.anthropic.com/v1/messages",
  "skills": ["~/.tapir/agent/skills", "/project/extra-skills"]
}

All fields are optional. Missing values fall back to environment variables, then to defaults.

Custom models

Register models with context window and cost info via the _models field:

{
  "_models": {
    "claude-sonnet-4-6": {
      "context": 200000,
      "max_output": 16384,
      "input_cost_per_m": 3.0,
      "output_cost_per_m": 15.0,
      "extended_thinking": true
    }
  }
}

System prompts

Tapir assembles the system prompt from multiple sources, in order:

1. Base prompt — .tapir/SYSTEM.md in the working directory, or ~/.tapir/agent/SYSTEM.md globally
2. Append — .tapir/APPEND_SYSTEM.md (project) and ~/.tapir/agent/APPEND_SYSTEM.md (global)
3. Context files — AGENTS.md or CLAUDE.md discovered from the home directory down to the working directory
4. Skills — loaded from ~/.tapir/agent/skills/, .agents/skills/, .tapir/skills/, and paths listed in config.json

Directory layout

~/.tapir/
  config.json          # global configuration
  history              # command history
  sessions/            # persisted sessions
  agent/
    SYSTEM.md          # global system prompt
    APPEND_SYSTEM.md   # appended to system prompt
    AGENTS.md          # global context
    skills/            # global skills

<project>/
  .tapir/
    SYSTEM.md          # project system prompt
    APPEND_SYSTEM.md   # project append
    skills/            # project skills
  .agents/
    skills/            # alternative skills location
  AGENTS.md            # project context

Running in a container

Since Tapir executes tools without confirmation prompts, running it inside a container keeps your host safe. Here is a minimal example using Alpine Linux:

Dockerfile

FROM alpine:latest

RUN apk add --no-cache git curl bash coreutils findutils grep

# Copy the static musl binary
COPY tapir /usr/local/bin/tapir
RUN chmod +x /usr/local/bin/tapir

# Create a non-root user
RUN adduser -D dev
USER dev
WORKDIR /home/dev/project

Build and run

# Build the image
docker build -t tapir-sandbox .

# Run with your API key and project mounted
docker run -it --rm \
  -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
  -v "$PWD":/home/dev/project \
  tapir-sandbox tapir

The -v flag mounts your project directory into the container. Tapir can read and modify those files, but cannot touch anything else on your host. When you are done, the container is discarded (--rm).

Tips

• Use the static musl binary — it runs on Alpine without extra libraries
• Add project-specific tools to the image (e.g. apk add python3 nodejs go) so the agent can build and test your code
• Mount ~/.tapir as a volume if you want to persist sessions and history across runs
• For extra isolation, drop capabilities with --cap-drop=ALL and disable networking with --network=none when the agent does not need internet access. The next release will support Ollama as a provider — you will be able to restrict the container's network to only reach your local Ollama instance, keeping the agent fully air-gapped from the internet

Commands

Slash commands

Shell escapes

Keyboard shortcuts

FAQ

Does Tapir support MCP?

No, and that is by design. The Model Context Protocol injects tool descriptions into the system prompt on every request, which inflates token usage significantly — even when those tools are never called. Tapir ships a fixed, minimal set of tools (read, write, edit, bash, ls, find, grep) that cover the vast majority of coding tasks. No protocol negotiation, no tool registries, no surprise token bills.

Is there a plan mode?

No. Just ask the agent to make one:

> Read the codebase and write a detailed implementation plan to PLAN.md

The agent will explore the project, reason about it, and save the plan to a file you can review, edit, and feed back in a later session. A plain markdown file is more portable and more useful than a built-in mode you cannot version-control.

Is there a todo list?

No. Same idea — ask the agent:

> Read TODO.md, pick the next item, implement it, and update the list

A TODO.md in your repo is grep-able, diff-able, and visible in every tool you already use. No proprietary format, no hidden state.

Screenshot

Download

v0.1.0

Linux x86_64 binaries:

• tapir-0.1.0-x86_64-linux-musl.tar.gz — static (musl libc)
• tapir-0.1.0-x86_64-linux-musl.zip
• tapir-0.1.0-x86_64-linux-gnu.tar.gz — dynamically linked (glibc)
• tapir-0.1.0-x86_64-linux-gnu.zip

Packages:

• tapir_0.1.0-1_amd64.deb — Debian/Ubuntu
• tapir-0.1.0-1-x86_64.pkg.tar.zst — Arch Linux

Source code:

• tapir-0.1.0-source.tar.gz
• tapir-0.1.0-source.zip

Verification:

• SHA256
• SHA256.sig
• tapir-010.pub — signify public key

signify -V -p tapir-010.pub -m SHA256 -x SHA256.sig
sha256sum -c SHA256

Source Code

• SourceHut (primary)
• GitHub (mirror)

License

ISC License

Copyright (c) 2025-2026 Murilo Ijanc <murilo@ijanc.org>

Permission to use, copy, modify, and/or distribute this software for any
purpose with or without fee is hereby granted, provided that the above
copyright notice and this permission notice appear in all copies.

THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.