nevercode

A fast, unrestricted AI coding agent written in pure C.

No telemetry. No bloat. No restrictions. Full freedom.

Features

• 58 built-in tools — shell, file I/O, search, git, web fetch, multi-agent delegation, MCP, and more
• Multi-provider — OpenAI-compatible (Chat Completions) and Anthropic (Messages API) with automatic format adaptation
• Long-running harness — 4-level context compaction, goal tracking, stall detection, checkpoint discipline, context handoff
• Multi-agent — parallel sub-agents, coordinator mode, team management, inter-agent messaging
• Session persistence — full replay from JSONL logs, cross-session memory, resume with context continuity
• Security-friendly — no command restrictions, built-in recon/exploit skills, CTF-ready
• ~79K lines of C, <10ms startup, <5MB memory, ~2MB binary, zero dependencies beyond libcurl

Build

make
make install          # optional — copies to /usr/local/bin

Requires: C11 compiler, libcurl (system). Optional: readline for line editing.

Platform Support

Platform	Status
macOS	Fully supported (primary target). File watching via kqueue.
Linux	Fully supported. File watching via inotify.
Windows	Not supported. Contributions welcome.

Quick Start

# OpenAI-compatible provider (default)
export OPENAI_API_KEY='your-key'
export OPENAI_BASE_URL='https://your-provider.example'
./nevercode

# Anthropic native provider
export ANTHROPIC_API_KEY='your-key'
./nevercode -P native

# One-shot prompt
./nevercode -p "explain quicksort in 3 sentences"

# Local Ollama
OPENAI_BASE_URL=http://127.0.0.1:11434/v1 OPENAI_API_KEY=dummy \
  ./nevercode -m llama3

# Resume previous session
./nevercode -c           # most recent
./nevercode -r <id>      # specific session

Security: never commit API keys. Use environment variables or -k flag only.

Configuration

Loading priority (later wins): built-in defaults → config file → env vars → CLI flags

Config File

~/.agents/config.json:

{
  "provider": "compat",
  "model": "gpt-5.4",
  "base_url": "https://your-provider.example",
  "max_tokens": 128000,
  "custom_instructions": "Always respond in English",
  "default_effort": "high"
}

API keys are never loaded from config files.

CLI Options

Flag	Description
-P, --provider	compat (default) or native
-b, --base-url	Custom API base URL
-m, --model	Model name
-k, --key	API key
-p, --prompt	One-shot prompt (non-interactive)
-s, --system-prompt	Custom system prompt
-t, --max-tokens	Max output tokens
-e, --effort	Reasoning effort: none/minimal/low/medium/high/xhigh/max
-j, --json	Structured JSON output
-r, --resume	Resume session by ID
-c, --continue	Resume most recent session
-F, --profile	Load config profile
-a, --approval	Approval preset: read-only / default / full-access
-D, --dangerously-skip-permissions	Skip all permission prompts
-V, --version	Show version
-h, --help	Show help

Environment Variables

Variable	Description
OPENAI_API_KEY	Compat provider API key
OPENAI_BASE_URL	Compat provider base URL
ANTHROPIC_API_KEY	Native provider API key
ANTHROPIC_BASE_URL	Custom native provider endpoint

Tools

58 built-in tools with full-access by default:

Category	Tools
File I/O	Read, Write, Edit, MultiEdit, ApplyPatch, Glob, Grep, ListDir, FileSearch
Execution	Bash, REPL, Notebook
Agent	Agent, SpawnAgent, ListAgents, WaitAgent, CloseAgent, SendMessage, AgentControl
Task	Task, TodoWrite, Cron, TaskGet/List/Stop/Update/Output
Team	TeamCreate, TeamDelete, Consensus
Git	GitDiff
Web	WebFetch, WebSearch
MCP	MCP, McpAuth, ListMcpResources, ReadMcpResource
Mode	EnterPlanMode, ExitPlanMode, EnterExecuteMode, EnterPairMode, ResetMode
Utility	Config, Brief, Sleep, ImageView, Scaffold, Snip, ProcMonitor, ToolSearch, ToolSuggest, AskUser, RequestPerms, Batch, Workflow, StructuredOutput
Worktree	EnterWorktree, ExitWorktree

Built-in Skills

Skill	Description
verify	Adversarial testing of code changes
debug	Diagnostic-driven debugging
simplify	Code refactoring and simplification
stuck	Break out of loops with alternative approaches
batch	Apply transformations to multiple files
security-review	Security-focused code review
perf	Performance profiling and optimization
test	Generate tests for recent changes
remember	Save notes to AGENTS.md for future sessions
loop	Iterative execution until objective is met
recon	CTF reconnaissance and enumeration
exploit	Exploit development and testing

Long-Running Harness

nevercode is built around harness engineering — the model is one component; the rest is durable orchestration.

Context Management

• L1 Snip: Zero-cost marker replacement for stale tool results
• L2 MicroCompact: Cache-aware bulk clearance when cache TTL expires
• L3 Collapse: Selective message folding without LLM calls
• L4 AutoCompact: Full LLM-based conversation summary with circuit breaker protection

Autonomy Features

Feature	Description
Context Handoff	Structured state serialization when context approaches exhaustion
Goal Evolution	Stack-based goal tracker with automatic sub-goal splitting
Stall Detector	Sliding-window watchdog for loop/no-progress/timeout detection
Session Watchdog	Per-turn wall-clock timeout with HTTP streaming cutoff
Context Fatigue	EMA-based predictor for proactive compaction before quality degrades
Tool Verification	Post-execution scanner for silent tool failures
Output Evaluator	Heuristic scorer for goal alignment and completeness
Auto-Test	Self-verification loop matching edited sources to test files
Checkpoint Discipline	Automatic git checkpoints every N file-changing turns
Ghost Snapshots	Transparent per-turn git snapshots on detached refs
Resource Monitor	Local CPU/memory/disk monitoring (no data leaves the machine)

Failure Handling

• 5 error categories (transient, persistent, resource, semantic, fatal) with tailored retry strategies
• Adaptive rate control with sliding-window pressure tracking (up to 8x backoff)
• Circuit breaker on compaction (3 consecutive failures → stop)

Multi-Agent

• Coordinator Mode: goal decomposition → task delegation → progress monitoring → result merge
• Agent Roles: default, explorer, awaiter, reviewer, recon, exploit, verifier (custom roles via JSON)
• Swarm Backends: in-process, tmux, or external process (auto-detected with fallback)
• Mailbox: inter-agent messaging with sequence tracking and FIFO delivery
• Team Memory Sync: shared key-value store with built-in secret scanning

Session System

• Full session replay from JSONL event logs
• Cross-session memory with provenance citations
• Rollout truncation for bounded disk usage
• Session modes (default, coordinator, plan, execute, pair) persisted per session
• Auto session titles from local heuristics (no API call)
• Background memory consolidation during idle periods

Plugin System

Shared-object plugins (.so/.dylib) in ~/.agents/plugins/ with lifecycle hooks:

void  *plugin_init(void);
void   plugin_unload(void *ctx);
int    plugin_hook(int event, const char *data, char **response, void *ctx);
const char *plugin_name(void);
const char *plugin_version(void);

Architecture

nevercode/
├── include/          # 15 headers (1332 macros in defaults.h)
├── src/              # 355 source files across 13 modules
│   ├── core/         # Main loop, config, conversation, state machine
│   ├── api/          # Provider implementations, streaming, auth
│   ├── tools/        # 58 built-in tools
│   ├── context/      # Context management, 4-level compaction
│   ├── agents/       # Multi-agent orchestration
│   ├── memory/       # Session history, memory pipeline
│   ├── exec/         # Tool execution, process management
│   ├── harness/      # Long-running orchestration
│   ├── ui/           # Terminal, rendering, vim mode
│   ├── utils/        # Fuzzy matching, encoding, caching
│   ├── review/       # Code review, risk assessment
│   ├── system/       # Permissions, sandbox, model registry
│   └── state/        # Session persistence, rollout tracking
├── tests/            # 408 test files (407 unit suites + 1 integration)
├── scripts/          # Development utilities
└── third_party/      # cJSON (embedded)

Stats

Metric	Value
Source	~79K lines of C (15 headers + 355 sources)
Tests	407 unit suites + 1 integration
Binary	~2.1MB (~1.9MB stripped)
Startup	<10ms
Memory	<5MB
Dependencies	libcurl (system), cJSON (embedded)
Telemetry	None

Testing

make test              # 407 unit test suites
make test-all          # Unit + integration
make test-harness      # Build + test + binary validation
make test-harness-json # With JSON report output

Project Instructions

nevercode loads project instructions from AGENTS.md (and legacy CLAUDE.md) with override/local/base precedence. Searches project tree upward and user home (~/.agents/).

License

See LICENSE.