Abhed documentation
Abhed is a deep agent harness. It runs where your code is, against whichever model you point it at, and records everything it does.
Guide
Using Abhed day to day.
Getting startedAbhed is a single static binary with no runtime to install.ConfigurationAbhed reads `.abhed/config.json` from the workspace. `abhed init` writes a starter file; everything below is optional and has a default.Models and providersAbhed does not ship a model. It is model-agnostic by construction: the harness is the same whichever endpoint you point it at, and a better model makePermissions and safetyAn agent that can run shell commands in your repository needs a real answer to "what is it allowed to do". Abhed's answer is a policy engine that deciToolsFour routes, none of which needs a rebuild. Pick by where your tool already lives.SkillsA skill is a procedure Abhed should follow when a request matches it: how your team cuts a release, the checklist for a schema migration, the house foExtensionsAn extension changes how the agent behaves without forking it. Extensions are separate processes speaking line-delimited JSON on stdin and stdout, so MCPAbhed is an MCP client and gateway. A tool exposed by a Model Context Protocol server appears in the model's tool list like any built-in one.The SDKAbhed's internals live under `internal/`, which Go refuses to let another module import. The `sdk` package is the supported surface across that wall, AutomationThree ways to run Abhed without a person at the prompt.Sessions and auditAbhed's state is a chronological stream of events. The agent is a function from that history to an action, and the runtime is a function from an actioObservabilityEvery session is already an ordered log of events — that is the audit trail, and it is the source of truth. Observability is the same facts in the shaStructured outputA program embedding Abhed usually does not want prose. It wants a value: a verdict, a list of findings, a record it can store. `RunJSON` runs a promptParallel subagents and worktree isolationThe `task` tool runs one subagent and waits. The `tasks` tool runs several at once. The agent chooses it when subtasks are independent — investigate tScheduled runsA recurring run is the shape of a lot of real work: every weekday at nine, summarise what changed overnight; hourly, check the staging pods; Sunday ni
Architecture
How it is built, and why.
Abhed — Design PrinciplesStatus: Draft · 2026-09-02 · Owner: Yuvraj SinghAbhed — System ArchitectureStatus: Draft · 2026-09-02Abhed — Execution Isolation & Prompt-Injection DefenseStatus: Draft · 2026-09-02Abhed — Resource Requirements & SizingStatus: Draft · 2026-09-02Abhed — Build Plan, Team & SequencingStatus: Draft · 2026-09-02 **Evidence status: [E].** Research produced no verified claims on build cost, team, or timeline. Estimates below are engineAbhed — Tool ContractsStatus: Draft · 2026-09-02 **Evidence status: [E] engineering judgment**, informed by verified finding P2 (tool surface discipline: 15 tools → 2 movedAbhed — System Prompt & Memory DesignStatus: Draft · 2026-09-02 **Evidence status: [E]**, structured around verified findings P1 (harness dominance), P2 (attention budget), P4 (re-injecteAbhed — Evaluation HarnessStatus: Draft · 2026-09-02 **Evidence status:** design [E], motivated by verified findings P1 and P10.Abhed — CLI & Product SurfaceStatus: Draft · 2026-09-02 · Evidence status: [E]Abhed — Data Model & APIStatus: Draft · 2026-09-02 · Evidence status: [E], structured by P6 (event sourcing).Abhed — Model providers and sampling parametersStatus: 2026-09-06Abhed extensionsStatus: 2026-09-06Embedding AbhedStatus: 2026-09-06
Operations
Running it in a real environment.
Abhed — Air-Gapped Deployment & Egress BrokeringStatus: Draft · 2026-09-02 **Evidence status: [E] engineering judgment throughout.** The research pass produced **zero verified claims** on air-gappedAbhed — Enabling AuthenticationFour modes. Pick by how Abhed is exposed, not by how much security sounds good.Abhed — Clusters, Machines, and External RetrievalFour capabilities that let Abhed work on infrastructure rather than only on files. All are **off by default**: reaching a cluster, a VM, or a corpus iAbhed — Choosing a ModelMeasured on an Apple M3 Pro, 36 GB unified memory, Ollama 0.33.2, September 2026. Reproduce with `abhed-modelcmp`; the numbers below are from this macAbhed — OIDC Provider ConfigurationVerified against the real token and JWKS shapes each provider issues (`internal/auth/providers_test.go`). The differences below are the ones that actuAbhed — Red-Team Engagement Scope**Status: scoping document. The engagement itself has NOT been performed.**Abhed — SkillsA skill is a folder with a `SKILL.md`: frontmatter naming it and saying when to use it, then instructions. It captures procedural knowledge that does
Trust
What a security review will ask, answered from the code.
Backup and restoreHow the hosted console's data is backed up today, how to restore it, and the recovery point that gives you. Written against `deploy/run.sh`'s actual cData handlingWhat Abhed stores, where it lives, how long it lives, what an operator can do about it, and who else touches data for the hosted console.Incident responseA short, honest runbook. Zybuu is a one-person company — this is what happens when there is one person, not a security operations center, and it is wrSecurity postureA summary for a security reviewer. Every claim below names the file that makes it true — read the file if you need more than the summary.Security scansIndependent, third-party tooling run against this repository and its container image on 2026-09-14. Every finding below is listed with its severity, iVulnerability disclosure policyQuestionnaire-formatted version of [`SECURITY.md`](../../SECURITY.md). If the two ever disagree, `SECURITY.md` is the source of truth — this document
Benchmarks
Measured, reproducible, published whatever they say.
Method: a reproducible local benchmarkThis directory holds a small, honest benchmark comparing three coding-agent harnesses driven by the **same local model**, so what's being measured is Results: Abhed vs. aider vs. a bare agent loop**Model:** `gemma4:26b`, served locally by Ollama at `http://127.0.0.1:11434` (OpenAI-compatible endpoint), identical for all three systems. **Hardwar