Codelim
AI Coding Agent by FunPulse Studios
A local-first AI coding agent built for reliable software changes.
Codelim gives one Main Agent control of the task while specialists work inside bounded, verifiable subtasks. Use your own models, your own tools, and your own workspace.
user> Audit our JWT token validation in auth/jwt.go and add automated unit tests with benchmarks.
[MAIN AGENT] Analyzing goal. Creating subtask contract for Explorer & Reviewer specialists...
↳ Subtask #1 (explorer): Discover symbol hierarchy & verify caller sites
EVIDENCE [OBSERVED] Found ValidateToken(tokenStr string) (*Claims, error) at auth/jwt.go#L35
EVIDENCE [OBSERVED] 4 HTTP handlers invoke ValidateToken without expiry grace check.
↳ Subtask #2 (reviewer): Verify security constraints & privilege boundaries
DISPOSITION ACCEPTED: Cryptographic signing valid. Recommended adding token expiration jitter.
[MAIN AGENT] Executing surgical modification via Implementer and running verification pipeline...
▶ write(auth/jwt_test.go) [MUTATION ACCEPTED: 1,420 bytes]
▶ bash("go test -v -bench=. ./auth/...") [PASS: 8/8 Tests, 0 Allocations]
✔ COMPLETE: JWT validation hardened, test suite created, verified against local test runner.
Engineered for Controlled, Verifiable Agentic Coding
Most coding assistants suffer from hallucinated tool loops, capability stripping, and uncontrolled prompt bloat. Codelim solves this with four non-negotiable architectural guarantees.
Sovereign Main Agent
The Main Agent remains the authoritative owner of the user's objective. Specialists are bounded workers, not competing agents fighting for control of the session.
Typed Subtask Contracts
Delegation is explicit and bounded. Specialists receive typed contracts with strict turn and tool budgets, returning structured SubtaskResult envelopes with evidence lineage.
Non-Escalation Authority Sandbox
Workspace (.codelim/agents) and user (~/.codelim/agents) markdown definitions extend capabilities without ever exceeding runtime security boundaries.
Bounded Generation Safety
Detect and contain oversized or invalid mutations with bounded writes and automatic rollback recovery before syntax errors compound.
Why Codelim?
An agent runtime for software engineering — engineered from first principles for predictable, verifiable outcomes.
| Traditional coding agents | Codelim |
|---|---|
| Can wander indefinitely | Bounded execution budgets |
| Specialists can become uncontrolled | Typed subtasks |
| Custom agents may request excessive capabilities | Authority ceiling |
| Large generation can corrupt files | Generation safety + rollback |
| Tool behavior is difficult to inspect | Structured evidence |
| Vendor / model dependent | Model agnostic |
| One persona controls everything | Main Agent + bounded specialists |
Built-in Specialists & Custom Agent Extensions
Built-in specialists provide core capabilities out of the box. User and workspace agents are extensions that operate inside strict capability ceilings.
Explorer (AgentExplorer)
Fast read-only codebase navigation, semantic symbol search, and architectural mapping.
Authorized Capabilities
- AST Grep & Semantic Symbol Search
- Structural Outline, Type Hierarchy, Call Hierarchy
- Fast recursive globbing and regex grep
Epistemic Truth Protocol
- [OBSERVED]: Verified directly in active source lines
- [INFERRED]: Deductions logically derived from code
- [UNKNOWN]: Unproven assumptions explicitly flagged
- Anti-wandering target checklist with early stop gating
Bring Your Own Agents (BYOA)
Extend Codelim without modifying its core runtime.
Define project-specific specialists in .codelim/agents/ or user-wide specialists in ~/.codelim/agents/.
The Main Agent can delegate appropriate subtasks to them, while the runtime enforces their authority ceiling.
Deterministic Delegation Cycle
↓
Subtask Contract → Custom Specialist Sandbox
↓
Structured Evidence → Main Agent
How Codelim Solves Complex Problems
Guesses are separated from observations, bounded by authority, and subjected to verification.
Objective Reception & Sovereign Main Loop
The user provides an objective. The Main Agent inspects workspace context with full tool sovereignty and plans execution.
Structured Subtask Contract Dispatch
When deep discovery or review is required, the Main Agent delegates a typed Subtask with explicit context, constraints, and budget.
Multi-Source Agent Catalog & Authority Enforcement
The catalog resolves the candidate agent across Workspace (.codelim/agents), User (~/.codelim/agents), and Built-in definitions, strictly clamping authority.
Evidence Envelope Return & Sovereign Continuation
The specialist returns a typed SubtaskResult with evidence and file diffs. The Main Agent synthesizes findings, executes verified mutations, and finishes the task.
Your Models. Your Machine. Zero Vendor Lock-in.
Codelim works with cloud and local models. Your code never leaves your workstation unless you explicitly send requests to your chosen provider.
Get Started in 30 Seconds
Single standalone binary with zero runtime dependencies. Available for Windows, macOS, and Linux.
macOS / Linux
Universal installer for ARM64 & x86_64
curl -fsSL https://funpulsestudios.com/codelim/install.sh | bash
Universal binary for macOS & Linux
Windows (PowerShell)
Native Windows x64 binary
iwr -useb https://funpulsestudios.com/codelim/install.ps1 | iex
Standalone executable with ANSI TUI
Everything You Need to Know
Fair questions. Straight answers. Everything about Codelim's architecture, security, and usage.
Codelim is a sovereign agent runtime, not just a chat interface. It acts as an autonomous pair programmer that reasons, inspects symbol hierarchies, executes builds and test suites, verifies outcomes, and rolls back failed changes autonomously.
Yes! Simply place markdown files in .codelim/agents/*.md (for project-specific specialists) or ~/.codelim/agents/*.md (for personal specialists). Codelim automatically discovers them and enforces strict security sandboxes.
When configured with local models (such as Ollama or vLLM), Codelim runs 100% offline with zero telemetry or network calls.
Codelim is developed and published by FunPulse Studios as part of our developer tools and real-time AI technology portfolio.
Build with an AI agent you can control.
Run Codelim locally, connect your preferred model, and give your engineering workflow bounded specialists, verifiable evidence, and runtime-enforced authority.