Daimonos README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Daimonos listing page.
Daimonos
An agent-optimized OS layer that makes AI coding agents faster and cheaper.
Daimonos replaces the built-in file, search, exec, and git tools in your AI coding agent with structured equivalents that return compact JSON instead of raw terminal output. The result: fewer tokens consumed, fewer round-trips, and lower API costs — with zero changes to your workflow.
Platforms: Linux (x86_64, aarch64) and macOS (Apple Silicon, Intel). Windows support is planned.
For repository agent/operator conventions, see AGENTS.md (especially
Daimonos tool usage policy).
The name comes from Greek daimon (agent/spirit), the etymological root of "daemon."
The problem
When an AI agent runs cargo test, it gets back hundreds of lines of terminal
output — progress bars, compile messages, passing test names — when all it
needs is "47 passed, 0 failed." The agent pays for every token of that noise:
reading it, reasoning about it, and carrying it in context for the rest of the
session.
The same waste happens with git status, docker ps, ls -la, and every
other shell command. Agents spend 30-50% of their token budget on verbose,
unstructured tool output.
How it works
Daimonos runs as an MCP server that your IDE or CLI spawns automatically. It provides the same operations agents already use — read files, write files, search, execute commands, git operations — but returns compact, structured JSON instead of raw text.
The single binary also provides ACP, one-shot agent, interactive chat, and socket-daemon runtimes — a full coding-agent harness in its own right; see Agent harness features below and Runtime modes for the explicit subcommands and compatibility aliases.
Four layers of optimization
-
Native tool plugins —
git,cargo,gh, anddockerare exposed as first-class MCP tools with structured JSON output. When agents callexec("cargo test"), Daimonos intercepts it and routes through the native plugin instead. -
Semantic output filters — For commands without native plugins (pytest, make, pip install, eslint, etc.), Daimonos applies semantic compression: test runners return summary + failures only, build commands return "ok" or just the errors, install commands return success/failure.
-
Protocol-level efficiency — Read deduplication (re-reading an unchanged file returns
{"unchanged":true}instead of the full content), compact field names, lazy tool exposure, batch operations, and a terse output directive that cuts LLM prose by ~30%. -
Managed subprocess execution — Command output is bounded while it is read instead of after full buffering. Daimonos owns Unix process groups, retires descendants on cancellation or session shutdown, isolates child environments through an explicit allowlist, and stores background output in private bounded artifacts.
Benchmark results
Tested with Claude Opus 4.6 on identical coding tasks (read files, search code, edit, run tests, git operations):
| Metric | Baseline | Daimonos | Savings |
|---|---|---|---|
| Output tokens | 5,842 | 3,198 | -45.3% |
| Total tokens | 41,239 | 33,847 | -17.9% |
| Tool calls | 17 avg | 14 avg | -17.6% |
| Wall time | 42.1s avg | 35.2s avg | -16.4% |
Remote benchmarks on AWS (same hardware, same model, same tasks) showed 20.3% cost reduction and 14.0% faster task completion.
SWE-bench Verified (mini) — three-way harness comparison
Five instances from swe-bench-verified-mini,
same model (Claude Opus 4.8) across all three harnesses, each agent running
inside the official SWE-bench Docker image for its instance (real test
environment), scored with the official swebench evaluation harness:
| Instance | daimonos tokens (LLM calls) | mini-swe-agent tokens (LLM calls) | cursor-agent tokens |
|---|---|---|---|
| django__django-11815 | 54,755 (4) | 38,305 (9) | 113,242 |
| django__django-12155 | 53,773 (4) | 33,095 (9) | 348,425 |
| django__django-12708 | 88,765 (6) | 196,492 (22) | 451,349 |
| sphinx-doc__sphinx-8035 | 172,729 (10) | 342,108 (30) | 930,495 |
| sphinx-doc__sphinx-9367 | 66,812 (5) | 18,611 (6) | 230,223 |
| Total tokens | 436,834 | 628,611 | 2,073,734 |
| Total wall time | 84 s | 247 s | 279 s |
| Resolved | 5/5 | 5/5 | 5/5 |
Conclusions:
- Correctness parity: all three harnesses resolved 5/5 at this sample size, so daimonos's token savings did not cost any resolutions.
- daimonos was cheapest and fastest: ~30% fewer tokens than mini-swe-agent (the minimal open-source baseline) and ~4.8x fewer than cursor-agent, with ~3x less wall time than either.
- Caveats: n=5, and run-to-run variance is real (daimonos spent 503k tokens on sphinx-8035 in an earlier identical-config run vs 173k here). cursor-agent's total is mostly cache-read tokens billed at a fraction of input price, so its raw token count overstates its relative cost. Measured API spend for the batch (OpenRouter): daimonos $1.53, mini-swe-agent $1.66; cursor-agent bills via Cursor's backend.
See benchmarks/swebench/ for the runners and methodology.
60-second demo
Use this script for README readers, release notes, and social posts:
What to highlight in the demo:
- same workflows, less tool-output noise
- structured responses instead of raw terminal spam
- fewer tokens and fewer round-trips for common coding tasks
Quick start
Install
Pre-built binaries (Linux and macOS):
From source:
See docs/install.md for all platforms (ARM Linux, Intel Mac, musl static builds).
Configure your IDE
For most users, start with one of these:
- Cursor: Cursor IDE setup
- Zed: Zed setup
- Claude Code: Claude Code setup
Add Daimonos as an MCP server. For Cursor, add to your project's
.cursor/mcp.json:
That's it. Daimonos starts when your IDE opens the project and exits when you close it. No daemon to manage, no background service.
Setup guides for other tools
- Cursor IDE
- GitHub Copilot (VS Code, Visual Studio, JetBrains, Xcode, Eclipse)
- Claude Code (CLI + macOS Desktop app)
- Windsurf
- Cline (VS Code extension)
- Gemini CLI
- Zed Editor
- Discord integration (bot token, allowlists, read-only tools)
- Other tools (Claude Desktop, ChatGPT, Continue.dev, BoltAI, etc.)
What's included
Core tools (always available)
| Tool | What it does |
|---|---|
read_file | Read with optional offset/limit, content-hash deduplication |
write_file | Write with auto-mkdir |
edit_file | String replacement with diff confirmation |
search | Regex search (content mode) or file discovery (file mode) |
exec | Run commands with semantic filtering, bounded capture, and owned teardown |
batch | Multiple operations in a single round-trip |
workspace_info | Project type, git status, directory listing, analytics |
Native tool plugins (auto-detected)
These appear automatically when the corresponding CLI tool is found on PATH:
| Plugin | Commands | Detected by |
|---|---|---|
git | status, log, diff, branch, add, commit, push, pull, checkout | .git directory |
cargo | test, build, check, clippy, fmt, add | Cargo.toml |
gh | pr_view, pr_list, pr_create, pr_diff, pr_checks, api | gh on PATH |
docker | ps, logs, exec, images, inspect, stop, compose_up/down/ps | docker on PATH |
Additional capabilities
- Workspace snapshots — Checkpoint before risky edits, rollback on failure
- Starlark scripting — Bundle multiple tool calls into a single script
- Token analytics — Per-tool-call tracking with cross-session history (
daimonos --stats) - Background processes — Start, poll, and stop long-running commands with admission limits, private bounded logs, and descendant cleanup
- Configurable — All tunables in a single TOML config file
Managed process lifecycle
Raw exec, background jobs, and CLI plugins (cargo, git, gh, docker,
npm, pytest, curl, and shellcheck) share one managed execution layer:
- Streaming-time bounds — stdout and stderr retain UTF-8-safe head/tail previews without first allocating the complete output
- Process-group ownership on Unix — cancellation and shutdown send TERM, wait a configurable grace period, then escalate to KILL and reap descendants
- Secure background artifacts — random exclusive
0600files under a private0700directory, with configurable byte and job-count limits - Environment isolation — children inherit only configured parent variables plus explicit session, tool, and per-call overrides; provider and MCP credentials are not ambiently leaked
- Structured-output integrity — plugins reject truncated JSON rather than reporting an incomplete result as valid
Agent harness features
Beyond the MCP server, the same binary is a complete coding-agent harness:
an interactive terminal UI (daimonos agent), an ACP backend for Zed, a
one-shot CLI, and a session daemon with attach/detach and remote control.
Many of its recent features come from a systematic study of 60+ open-source agent harnesses (Aider, OpenHands, SWE-agent, Goose, OpenCode, Forge, Pi, the Cline family, and others) — mining the ecosystem for proven techniques and adapting the best ones.
Provider resilience — a hiccup never kills the turn
- Bounded provider retries with backoff for transient failures (429/5xx/ network), classified at the provider boundary — fatal auth/validation errors surface immediately
- Automatic model failover — on a sustained overload the turn continues on the next model in the chain, then returns to your preferred model on the next turn
- Turn-level error resume — when retries and failover are spent, the agent pauses, repairs the conversation (keeping partial streamed output), and continues where it left off; recovery actions surface in the UI
- Retry-storm detection — fingerprints repeated identical tool calls and steers the model out of loops
- Orphan tool-call repair — max-token truncation mid-tool-call is repaired instead of poisoning the session
Multi-provider sessions
- Several providers, one session — configure Anthropic, OpenAI, and
OpenRouter side by side (
DAIMONOS_AGENT_<NAME>_API_KEY); every call is routed to the right provider by model, with an explicitprovider:slugoverride - Live model discovery — at startup the configured provider(s) are queried for their full model catalogs; the model picker and failover chain always reflect what is actually served, newest first
- Cross-provider failover — with more than one provider configured, an outage at one can fail over to models at another, mid-turn
- Provider-reported context windows — compaction thresholds derive from the live model metadata instead of hardcoded numbers
Context economy at the harness level
- Conversation compaction — summarize-and-continue with high/low water-mark thresholds and provider-honest token accounting
- Bounded tool results — oversized tool output is capped at the dispatch boundary and offloaded to files the agent can re-read selectively
- Reverse-budget pruning — old tool results shrink before new ones, so a long turn keeps its recent working set sharp
- Distilled working memory — durable facts/snippets/notes that survive compaction, separate from the transcript
- Resilient edit matching (mined from Aider) — whitespace-tolerant search/replace cuts failed-edit retry costs
- Batched scripting — the agent is steered to bundle multi-step tool work into single Starlark scripts (~2.2x cost lever, benchmarked)
Session durability and control
- Per-turn workspace checkpoints — automatic snapshots with diff/compare and code-only restore
- Daemon-owned sessions — detach from a running agent, reattach later (or from another terminal), with a reconnect event ring and canonical snapshot recovery
- Persistent terminal UI — streaming output, tool-lifecycle cards, approval modal, model/usage status bar, and vim-style scrollback
- Remote control — paired Android controller over an authenticated WSS gateway
- Subagent delegation — drive external ACP agents (cursor-agent, codex-acp, …) as delegated workers
- Thought capture — opt-in local persistence of streamed model reasoning for later inspection
Agent-mode configuration lives in a dotenv-style agent.env
(~/.config/daimonos/agent.env); see Runtime modes.
Architecture
Daimonos is a single Rust binary with two planes that share one tool implementation, one opcode protocol, one config, and one analytics store:
- Tool server — speaks MCP over stdio (or a Unix socket) to an external agent. Your IDE spawns it as a subprocess — no network, no containers, no setup beyond a JSON config entry.
- Agent harness — runs the agent loop itself, dispatching those same tools in-process (no MCP hop) and talking to LLM providers directly.
Tool-server plane
Under the hood, Daimonos uses an opcode-based protocol where each operation
has a numeric identifier and compact field names (c, p, s, n) to
minimize token overhead. The MCP layer translates between standard JSON-RPC
and the internal opcode format.
Agent-harness plane
Every frontend drives the same transport-independent session core, so a
conversation started in the TUI can detach to the daemon and be reattached
from another terminal or a paired phone. Provider adapters own all
provider-specific wire format and error classification; everything above
them sees one LlmProvider interface and plain model strings — which is
what makes failover, live model discovery, and multi-provider routing
composable rather than special-cased.
Project vision
Daimonos is being built in three phases:
Phase 1: User-space MCP server + agent harness (current)
A Rust binary that runs on any Linux or macOS machine, in two roles that prove out the same protocol design and structured I/O patterns:
- Tool server for third-party agents (Cursor, Copilot, Claude Code, Zed, …) via MCP — the original phase-1 deliverable.
- Agent harness in its own right: interactive TUI, ACP backend for Zed, one-shot CLI, and daemon-owned sessions with remote attach — with multi-provider routing, model failover, compaction, and per-turn checkpoints built in (see Agent harness features).
Status: Production-ready. Both roles are used daily for real development work — including developing Daimonos itself. Pre-built binaries available for Linux (x86_64, aarch64, musl) and macOS (Apple Silicon, Intel).
Phase 2: Minimal Linux distro
A purpose-built Buildroot Linux image with Daimonos as the primary user-space application. Designed for cloud deployment where AI agents need a clean, minimal environment. The distro boots in seconds, has no shell or human-facing UI, and runs the Daimonos daemon as PID 1's direct child. The session daemon and remote-control gateway from phase 1 are the intended tenants: headless agent sessions in the cloud, attached to from a terminal or phone.
Status: Working prototype. Boots in QEMU, deployable to AWS EC2. Used for remote benchmarking.
Phase 3: Custom microkernel
The long-term vision: a microkernel where Daimonos opcodes become native syscalls. StructFS (a filesystem that stores and returns structured data natively), capability-based security, and a process model designed for agent workloads from the ground up.
Status: Design phase.
Development
Prerequisites
| Dependency | Required | Install |
|---|---|---|
| Rust (stable 1.75+) | Build | rustup.rs |
| Python 3 + pytest | Tests | pip install -r tests/requirements.txt |
Running tests
Running benchmarks
See benchmarks/README.md for details.
Configuration
All behavior is tunable via a TOML config file. See docs/configuration.md for the full reference, or daimonos.default.toml for annotated defaults.
Key sections:
[index]— Trigram indexer tuning (max depth, file size limits)[search]— Search result limits[process]— Process timeouts, in-memory/artifact bounds, background admission, termination grace, inherited environment, semantic filters, and max concurrent Starlark script threads[pipeline_cache]— Subprocess result cache size, inotify watch cap, extra ignored directories[analytics]— Token tracking (SQLite storage, retention)[tools.*]— Per-tool plugin configuration
Contributing
Daimonos is in active development. If you're interested in contributing, start with the AGENTS.md file for coding conventions, architecture decisions, and the review checklist.
See also:
License
MIT