A personal coding-agent harness in Go. One static binary. Instant to open.

Go 1.24+MIT LicenseStartup 5.9msBinary 19MBCGO Disabled

Paused

Bring your own key

12providers

Anthropic
OpenAI
Google
Ollama
OpenRouter
OpenCode Zen
Kimi Code
Kilo Code
NVIDIA
Groq
Together
DeepSeek

One pipeline, one config file. Route a different model to every stage.

Four agents turn a task into a branch

Planner, Coder, Tester, and Reviewer each start from a fresh context. They hand the next stage typed artifacts instead of one ever-growing conversation, so every step stays independently checkable.

How the pipeline works

niki run

Task inPlanner

TaskSpec

Planner

Reads the task and the current files, then defines the smallest safe plan.

Context: Fresh LLM session
Handoff: TaskSpec

Planner selected

Nothing lands until the tests pass

A run writes a fresh niki/<id> branch only when the diff is non-empty and the executed test suite passed. A failed suite blocks the branch unless you force the run, and committed history is never rewritten.

Read the security model

niki status

mainniki/4f2a

niki run "add a /health endpoint"

Non-empty diffrequired
Executed test suiterequired
niki/<id> createdrequired
Committed historynever rewritten

Branch niki/4f2a · 1 commit · main untouched

The run leaves receipts you can read

Every run writes changes.patch, report.md, and one artifacts/*.json per agent. Inspect them before you decide what belongs in your branch.

curl -fsSL https://raw.githubusercontent.com/RavaniRoshan/niki/master/scripts/install.sh | bash

Read the CLI reference

niki run

plan.mdPlan mode

Explicit plan mode researches without executing. Approve the plan with niki run --plan <id>.

# Add a /health endpoint
 
## Smallest safe change
- register GET /health in the existing router
- reuse the db ping readiness already makes
- one test, asserting 200 and the body shape

Sample output. The real file ships with the run and stays with the branch.

Route a different model to every stage

Put a stronger reasoner on planning and review, a faster model on testing, or mix providers per stage. Niki is bring-your-own-key end to end.

See provider setup

niki.tomlreading config


[general]
max_revision_rounds = 3
spend_cap_usd = 5.0
[agents.planner]
provider = "anthropic"
model = "claude-sonnet-4-20250514"

[agents.coder]
provider = "anthropic"
model = "claude-sonnet-4-20250514"

[agents.tester]
provider = "openai"
model = "gpt-4o-mini"

[agents.reviewer]
provider = "anthropic"
model = "claude-sonnet-4-20250514"

Performance contracts. No feelings, just measurements.

Every number is real, measured on host with the included PTY probe harness and hyperfine. NIKI boots to first frame 3.96x faster than Codex, consumes 39% less memory, and compiles to a binary 12.8x smaller.

ToolVersionLanguage--versionTTFPIdle MemoryBinary

Codex CLI0.152.1Rust20.0 ms23.4 ms21.7 MB244 MB

Google agy1.3.1Go19.3 ms778.5 ms225.5 MB202 MB

Kimi Code2.1.1TS/Node188.4 ms1253.7 ms391.6 MB75 MB

NIKI0.11.0Go6.8 ms5.9 ms13.1 MB19 MB

Built-in tool suite. Zero fluff.

Seven core coding tools, each with explicit safety guarantees. No bloat, no framework sprawl, just what a coding agent needs.

read_fileRead file contents with line rangesBounded size limits, path sanitization

write_fileAtomic write file replacementTemporary file swap, prevents corruption

edit_fileTargeted string replacementRequires exact match, rejects ambiguous edits

apply_patchUnified diff patch applicatorFuzzed parser (>400k iterations with 0 panics)

globFind files matching wildcard patternsBounded directory walking, ignore awareness

grepFast regex content searchRead-only concurrency, skip binary files

shellExecute shell commands in workspaceBubblewrap isolation, network denial, dropped caps

Every run is auditable.

Three files stand between a finished run and your merge decision. No dashboards to trust, nothing inferred on a server you cannot read.

changes.patch

The exact diff

--- a/src/server.ts
+++ b/src/server.ts
@@ -41,6 +41,12 @@
+router.get('/health', async (_req, res) => {
+  const db = await ping();
+  res.status(db ? 200 : 503).json({ ok: db });
+});

A unified diff written from the completed run and bound to the branch when branch creation succeeds.

report.md
What happened, in order
- Planner3 steps
- Coder+12 -6
- Tester4 passed
- Reviewerapproved
A readable account of every stage, its verdict, and the reason the run stopped or continued.
artifacts/*.json
Per-agent decisions
```
{
  "agent": "reviewer",
  "decision": "approve",
  "revisions": 0,
  "blocking": true
}
```
One JSON file per agent recording what it decided and why, so a reviewer can argue with the call.

Changelog

2026-10-08v0.11.0TUI redesign, Go rewrite, and platform hardening
2026-09-23v0.8.0Agent-harness + runtime release
2026-09-08v0.7.0Mega-plan execution
2026-08-21v0.6.0Claude Code parity, interaction, trust, and ecosystem surface

See what's new in Niki

What you can change

Pick the agents
Four named stages run by default, and specialist agents can be added behind risk gates. Explicit pipeline topologies are never rewritten.
Explore agents
1. PlannerTaskSpec
2. Coderunified diff
3. Testertest results
4. Reviewerverdict
Choose the sandbox
Podman or Docker for isolation, or a git worktree backend that needs no container runtime at all.
Explore integrations
- PodmanDefault rootless container sandbox with no daemon.
- DockerContainer sandbox fallback. Docker writes through the project mount.
- worktreeGit worktree backend that needs no container runtime.
Set the guardrails
Spend caps, a single hysteresis budget across retries, and command deny lists, all in one config file.
Explore security
- spend_cap_usdaborts the run
- max_stepsaborts the run
- max_wallclock_secsaborts the run
- command deny listfails closed

Niki is free software you can read end to end.

MIT licensed, no telemetry, your own provider keys. Every prompt, every gate, and every artifact is in a repository you can fork tonight.

Read the source

Runs on your machine

PodmanDefault rootless container sandbox with no daemon.
DockerContainer sandbox fallback. Docker writes through the project mount.
worktreeGit worktree backend that needs no container runtime.

Talks to 12 providers

Anthropic, OpenAI, Google, Ollama, OpenRouter, OpenCode Zen, Kimi Code, Kilo Code, NVIDIA, Groq, Together, DeepSeek.

Rule of ProofWorks means a real test or measurement ran and its output was verified.

Zero Proprietary CodeNo leaked, decompiled, or reconstructed proprietary source was ever accessed or copied.

Small and BoringDependencies pass strict admission checks. Zero framework sprawl.

Recent highlights

Read the documentation

One sentence in. A reviewable branch out.

Download Niki