CodingAgent
■ SOVEREIGN · LOCAL-LLM-FIRST · MCP-NATIVE
Sovereign, open-source, local-LLM-first agentic engineering for planning, coding, debugging, review, security and verification.
Models reason, tools act through explicit boundaries, evidence records what happened, and verification determines whether work is actually done.
01 // INFERENCE
LOCAL-FIRST
Ollama, vLLM, llama.cpp and LM Studio as first-class targets for private repos.
02 // PROTOCOL
MCP-NATIVE
Governed tool discovery and schema contracts.
03 // EVIDENCE
TEST-LED
Compiler exit codes determine completion.
04 // POLICY
STRICT GATE
Explicit human approval on write and network actions.
■ How it works
Every task runs the same deterministic loop. The model never gets to declare its own work finished — verification does.
1. Model reasons
A local LLM plans the task and proposes a course of action.
2. Tools act
Execution happens through MCP-governed, schema-validated tool calls.
3. Evidence records
Every action is logged — what ran, what changed, what it returned.
4. Verification decides
Compiler exit codes and test results — not the model — determine completion.
■ Agent modes
Each mode carries its own tool allowlist, restricted actions and verification requirement — not just a different prompt.
Decompose complex goals, map dependencies, surface uncertainty and define acceptance criteria before code changes begin.
VERIFIES: Plan DAG must satisfy schema validation, contain no circular dependencies, and define unit tests for every executable task unit.
Implement scoped changes across files while preserving repository conventions and explicit tool boundaries.
VERIFIES: Every modified file must compile cleanly, pass syntax linter checks, and preserve code style guidelines.
Trace failures through logs, tests and code paths, then produce evidence-backed hypotheses and fixes.
VERIFIES: A reproducing unit test must fail before the fix and pass unconditionally after applying the patch.
Evaluate diffs for correctness, maintainability, regressions and missing tests without self-certifying the authoring agent.
VERIFIES: Zero unaddressed critical severity flags and confirmed test coverage over newly added code branches.
Inspect trust boundaries, unsafe tool use, injection paths, secrets exposure and dependency risk.
VERIFIES: Cryptographic attestation certifying zero critical CVEs and zero high-entropy secret leaks in staged artifacts.
Explore a codebase or architecture without changing it—useful for onboarding, design review and repository understanding.
VERIFIES: Every generated explanation must reference verified, existing repository file paths and symbol names.
■ Local runtimes
Private repositories route strictly to local models, with hardware-aware dispatch across VRAM, context size and verification history.
LOCAL
Instant local model serving with automatic GPU acceleration across Metal, CUDA, and ROCm. Perfect for single-developer workstations.
LOCAL
Continuous batching and PagedAttention for serving multiple concurrent agent subroutines across shared on-premise GPU clusters.
LOCAL
Lightweight C/C++ engine executing quantized GGUF weights with zero dependencies and efficient CPU/RAM fallback.
LOCAL
Visual desktop application providing interactive model experimentation, hardware utilization monitoring, and local HTTP endpoint.
■ Security posture
9 capability rules across filesystem, network, git, deploy, database, shell and secrets — each with an explicit posture and mitigation, not a blanket policy.
Path normalization prevents traversal outside task workspace.
Restricted strictly to active git worktree; cannot touch root dotfiles or .git/config.
Domain allowlisting + operator confirmation dialog specifying destination URL.
Operator must visually review full unified diff and sign commit approval.
■ The knowledge graph
Agents, MCP, local runtimes, security, verification, enterprise tooling and India-first solutions — each pillar with explicit rationale and a verification aspect.
CodingAgent Core & Agents
10 pillars
CodingAgent MCP & Integrations
10 pillars
CodingAgent Local LLMs & Routing
10 pillars
CodingAgent Security & Sovereignty
10 pillars
CodingAgent Verification & Memory
10 pillars
CodingAgent Enterprise & Tooling
10 pillars
CodingAgent India & Solutions
10 pillars
Platform Pillars
10 pillars
Local inference, governed tools, verified outcomes. Open source, MIT-licensed.